Methods, devices, equipment and media for copyright risk control in generative content reasoning process
By unifying the representation and dynamically adjusting copyright risk factors during generative content reasoning, the problem of incomplete copyright risk control in existing technologies is solved. This enables proactive avoidance of high-risk expressions during the generation process, reducing computational costs and the probability of similar outputs, and effectively controlling copyright risks while maintaining the user's creative intent.
Patent Information
- Application Number
- CN202611239912.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-17
AI Technical Summary
Existing generative content systems lack a unified risk control mechanism, making it difficult to uniformly represent and dynamically adjust copyright risk factors scattered across text, images, videos, audio, and generation parameters before and during inference. This results in generated results that may be highly similar to protected expressions, and also incurs high computational costs.
By acquiring the generation conditions, preprocessing is performed to generate target feature vectors, copyright risk factors are identified and encoded as copyright risk vectors, copyright risk control instructions are generated, prompt words, attention weights, sampling trajectories and negative constraints are adjusted during the generation process, and similarity comparison is performed to ensure that the generated results are far away from high-risk protected expressions.
It enables proactive avoidance of high-risk protected expressions during the generation process, reduces the computational cost of repeated retries after generation and the probability of highly similar outputs, and effectively controls copyright risks while maintaining the user's creative intent.
Smart Images

Figure CN122818317B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to methods, apparatus, equipment and media for copyright risk control in generative content reasoning processes. Background Technology
[0002] Generative AI models can generate images, videos, music, and text based on user-input prompts, reference images, reference videos, audio clips, or multimodal conditions. With the development of diffusion models, video generation models, and large multimodal models, users can generate highly specific content with minimal text or reference materials. For example, users can input descriptions such as "a yellow electric animal representing a classic animated character," "a neon-lit street scene in a rainy night from a famous movie," or "the drumbeats and melody of an early work by a singer." Even without directly using the work's title, the model may generate content highly similar to protected works. If the generated result is highly similar to existing copyrighted works, character images, brand logos, portraits, film and television scenes, works of art, or musical works, it will pose significant copyright risks to content platforms, model service providers, and users. To control these risks, existing generative content systems typically only implement isolated copyright risk control at the input or output end, mainly including the following implementation schemes. The first type is the input-end rule review scheme. This type of solution matches user prompts with keywords, entity names, brand names, or sensitive phrases. If a blacklist is matched, generation is refused or the user is prompted to modify it. Some generative content products also prohibit users from inputting prompts that infringe on third-party copyrights, trademarks, or other rights, or from uploading reference materials containing protected third-party content, through user guidelines and other compliance rules. This type of solution is simple to implement, but it can only handle explicit word risks or rely on user self-discipline, and it is difficult to identify implicit anchoring risks formed through indirect descriptions, reference images, reference videos, or music clips. The second type is the prompt security rewriting solution. When this type of solution detects the presence of protected entities or high-risk expressions in the prompts, it uses rules or language models to rewrite the prompts into more abstract descriptions, such as replacing specific character names with general character categories, or replacing specific artist names with general style descriptions. Related research has also proposed a copyright shielding framework that combines detection modules, prompt rewriting modules, and guidance adjustment modules during inference to adjust the impact of original prompts and security prompts on the generation process. This type of scheme can reduce some input risks, but it usually only processes the prompt text itself, without incorporating reference images, video frames, audio rhythms, generation parameters, and intermediate inference states into the same control loop. The third type is post-generation similarity filtering. This type of scheme first calls the generation model to obtain candidate generation results, and then compares the candidate generation results with a protected content library for similarity. If the similarity exceeds a threshold, the candidate generation result is discarded and regenerated, or the random seed is adjusted, the prompt words are slightly modified, and then it is generated again. Its disadvantage is that high-risk content has already been generated, resulting in wasted computation; if multiple retries still attract the same high-risk semantics, the generation efficiency and usability are low. The fourth type is model security alignment or training data governance schemes. This type of scheme reduces the probability of the model reproducing protected content through model training, fine-tuning, data filtering, or watermarking mechanisms.This type of solution is effective for model service providers, but it is difficult to quickly apply to already deployed third-party generated models or black-box models, and it is difficult to perform fine-grained risk control for individual user tasks.
[0003] In summary, existing solutions typically lack a unified risk vector mechanism that works both before and during inference. They cannot separate the retainable portion of user intent from the risky portion that might induce protected expression, and further convert this into executable model inference control instructions. How to uniformly represent the copyright risk factors scattered across text, images, videos, audio, and generation parameters into a computable copyright risk vector before and during generative content inference, and dynamically adjust the sampling trajectory, cue word expression, guiding weights, and negative constraints of the generation model based on this copyright risk vector, so that the generated result actively avoids high-risk protected expression while maintaining the user's creative intent, is a pressing technical problem in this field. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for copyright risk control in generative content reasoning. This method can uniformly represent copyright risk factors scattered across text, images, videos, audio, and generation parameters as a calculable copyright risk vector before and during generative content reasoning. Based on this copyright risk vector, the generation process is dynamically adjusted so that the generated result actively avoids high-risk protected expressions while maintaining the user's creative intent, reducing the computational cost of repeated retries after generation and the probability of highly similar outputs. The specific solution is as follows: Firstly, this application discloses a method for controlling copyright risks in the generative content reasoning process, including: The generation conditions are obtained and preprocessed to generate a target feature vector. Based on a preset protected content library and a preset copyright risk feature library, each copyright risk factor in the target feature vector is identified. Each of the copyright risk factors is encoded to obtain a copyright risk vector, a comprehensive risk score corresponding to the copyright risk vector is determined, and a copyright risk control instruction is generated based on the comprehensive risk score and the copyright risk vector. During the generative content reasoning process, the generation process is controlled according to the copyright risk control instructions to obtain the corresponding candidate generation results; the control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints. The candidate generated results are compared for similarity based on a preset protected content library and a preset copyright risk feature library. A target risk score is determined based on the comparison results. Based on the target risk score, a target generated result that meets the preset copyright risk threshold is determined, and the corresponding copyright risk control record is output.
[0005] Optionally, the generation conditions include any one or a combination of several of the following: prompt words, reference images, reference videos, style words, text semantics, shot descriptions, music clips, audio rhythms, seed values, resolution, frame rates, generation duration, sampling steps, guidance intensity, model type, and risk control intensity selected by the user. Accordingly, the preprocessing of the generation conditions to generate the target feature vector includes: The text, images, videos, audio, and parameters are standardized respectively to obtain processed data; the processed data includes any one or a combination of several of the following: vector representation of text semantics, vector representation of image vision, vector representation of video frame sequence, vector representation of audio spectrum, rhythm features, melody contour features, and generation parameter features. The processed data is concatenated or fused to obtain the target feature vector; wherein, if the processed data contains only one vector representation, then the vector representation is directly determined as the target feature vector.
[0006] Optionally, the copyright risk factors include any one or a combination of several of the following: protected entity risk factors, style risk factors, composition risk factors, shot risk factors, text semantic risk factors, audio rhythm risk factors, melody risk factors, reference material similarity risk factors, and generation parameter risk factors. The generation parameter risk factors are determined based on the association rules between at least one of the generation parameters (seed value, resolution, frame rate, generation duration, sampling steps, model type, and guidance intensity) and the protected content in the generation conditions. The preset protected content library is a library storing protected content objects. The preset copyright risk feature library is a library storing machine-processable features calculated using the preset protected content library. The machine-processable features include any one or a combination of several of the following: text semantic vector representation, entity alias list, global image vector representation, local visual fingerprint, color histogram, protected representation diagram, composition relationship diagram, video shot sequence features, motion trajectory features, audio fingerprint, rhythm features, melody outline, and style tags. Each machine-processable feature in the preset copyright risk feature library is associated with the corresponding protected content object in the preset protected content library through a protected content identifier. Accordingly, the step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library includes: Determine the features to be identified from the target feature vector, corresponding to at least one modality among text, image, video, and audio; Based on the machine-processable features in the preset copyright risk feature library, at least one of entity matching, alias matching, vector recall, structure matching and combination rule calculation is performed on the feature to be identified to determine the protected content object hit by the feature to be identified, and the first similarity between the feature to be identified and the hit protected content object. Based on the first similarity and the risk level, work type and ownership information of the matched protected content object recorded in the preset copyright risk feature library, generate each copyright risk factor of the matched protected content object; The record information corresponding to each copyright risk factor includes any one or a combination of several of the following: risk factor identifier, corresponding generation task identifier, risk source modality, risk type, source location, standardized risk expression, associated protected content identifier, first similarity, identification confidence level, risk severity level, and suggested avoidance action; the risk source modality includes any one of text, image, video, audio, and parameters; the risk type includes any one of protected entity, style, composition, shot, text semantics, audio rhythm, melody, similarity of reference materials, and generation parameters.
[0007] Optionally, the step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library further includes: At least two of the following characteristics corresponding to the target feature vector—color, shape, clothing, props, actions, scenes, shots, timbre, and rhythm—are used as nodes, and at least one of the following characteristics—the same subject, the same shot, consecutive frames, and the same audio window—are used as edges between the nodes to construct an attribute combination graph. The attribute combination graph is matched with the protected expression graph in the preset copyright risk feature library by subgraph matching or graph vector representation similarity comparison. When the similarity between multiple nodes in the attribute combination graph and the protected expression graph in terms of combination relationship meets the preset conditions, a combination risk factor is generated. The context enhancement coefficient of the combined risk factor is determined based on at least one of the following: number of hit nodes, node rarity, relationship matching degree, and source aggregation degree. Copyright risk factors identified from different modalities are mapped to the same protected content identifier. Cross-modal consistency is determined based on the ratio of the number of modalities that hit the same protected content identifier to the total number of modalities involved in the generation conditions. The context enhancement coefficient of the corresponding copyright risk factor is adjusted based on the cross-modal consistency.
[0008] Optionally, encoding each of the copyright risk factors to obtain a copyright risk vector and determining the comprehensive risk score corresponding to the copyright risk vector includes: The copyright risk factors are mapped to a unified risk space to obtain the copyright risk vector; the copyright risk vector includes at least two risk components from the following: protected entity risk component, style risk component, composition risk component, shot risk component, text semantic risk component, audio rhythm risk component, melody risk component, reference material similarity risk component, and generation parameter risk component. The risk components are determined based on the weights of the target risk dimensions corresponding to each risk component, the second similarity between the target feature vector and the preset protected content library on the target risk dimension, the identification confidence, the source reliability, and the context enhancement coefficient; wherein, the source reliability is determined based on the risk source modality of the copyright risk factor, and when there are multiple copyright risk factors in the same risk dimension, the risk component corresponding to the risk dimension is determined based on the maximum value, weighted average, or Top-K weighted aggregation method; Each of the aforementioned risk components is weighted according to the first-order weight of the corresponding target risk dimension to obtain the first-order weighted result; The product of any two different risk components is weighted according to the corresponding interaction weights to obtain the interaction weighted result. The summation of the first-order weighted result and the interactive weighted result is normalized to obtain the comprehensive risk score.
[0009] Optionally, generating copyright risk control instructions based on the comprehensive risk score and the copyright risk vector includes: The copyright risk vector is mapped to a copyright risk control instruction comprising at least two of the following: a mixed weight of the vector representation of the original prompt word and the vector representation of the safety prompt word; a generative model guidance strength; a negative constraint weight; a sampling trajectory offset strength; a safety prompt word; a set of negative constraints; a similarity verification threshold; and the number of local resampling times. Specifically, if the overall risk score is less than a first preset threshold, the original prompt word and reference material are retained; if the overall risk score is greater than or equal to the first preset threshold but less than a second preset threshold, the safety prompt word is generated and the negative constraint weight is increased; if the overall risk score is greater than or equal to the second preset threshold, the influence of the vector representation corresponding to the high-risk segment in the original prompt word is reduced, the influence of the vector representation of the safety prompt word and the negative constraint weight are increased, and the intermediate state of inference is reviewed. Accordingly, generating the security prompt word includes: The original prompt words are segmented into multiple semantic segments, and each semantic segment is aligned with the copyright risk factor and marked as a segment that can be retained, a segment that needs to be weakened, a segment that needs to be replaced, or a segment that needs to be deleted. The fragments to be weakened are abstracted to obtain an abstract description; Generate alternative expressions for the segments that need to be replaced, and combine the retainable segments, the alternative expressions, and the abstract descriptions into the security prompt words; Determine the semantic similarity between the security prompt word and the original prompt word after removing the weakened, replaced, or deleted segments; If the semantic similarity is less than a third preset threshold, the rewriting intensity is reduced; wherein the rewriting intensity is positively correlated with the difference between the security prompt word and the original prompt word.
[0010] Optionally, the step of controlling the generation process according to the copyright risk control instruction to obtain the corresponding candidate generation results includes: In the generative content reasoning process, intermediate inference states are extracted based on a preset step size or event triggering method; the intermediate inference states include at least one of intermediate latent variables, attention maps, low-resolution preview features of the current predicted image, video keyframe features, and audio segment features; The intermediate reasoning state is compared with the preset copyright risk feature database to obtain the current intermediate risk score; Based on the relationship between the current intermediate risk score and the previous intermediate risk score corresponding to the previous review step, the sampling trajectory offset intensity corresponding to the sampling trajectory adjustment, the constraint intensity corresponding to the negative constraint enhancement, and the attention weight corresponding to the high-risk segment in the original prompt word are updated to obtain the updated copyright risk control instruction; the sampling trajectory offset intensity and the constraint intensity are parameters based on the copyright risk control instruction. According to the updated copyright risk control instructions, the process jumps back to the step of extracting intermediate states of inference based on preset step size or event triggering method until the inference is completed, so as to obtain the corresponding candidate generation results.
[0011] Optionally, the step of comparing the candidate generated results based on a preset protected content library and a preset copyright risk feature library, determining a target risk score based on the comparison results, and determining a target generated result that meets a preset copyright risk threshold based on the target risk score includes: Based on the preset protected content library and the preset copyright risk feature library, the candidate generated results are compared with at least two of the following similarities: visual similarity, text semantic similarity, audio similarity, structural similarity, and entity recognition similarity, to obtain the corresponding comparison results. The similarity scores in the comparison results are weighted and summed based on the corresponding preset weights to obtain the target risk score; If the target risk score is greater than or equal to the preset copyright risk threshold, then the risk location result in the candidate generation result is determined according to the comparison result; the risk location result includes any one or a combination of several of the following: local image regions in an image, shot segments in a video, and audio windows in an audio file. Based on the risk location results, local redrawing is performed on the corresponding local image regions, segment resampling is performed on the corresponding video shot segments, or section replacement is performed on the corresponding audio windows to obtain the target generation result that meets the preset copyright risk threshold.
[0012] Secondly, this application discloses a copyright risk control device in the generative content reasoning process, comprising: The copyright risk factor identification module is used to obtain the generation conditions, preprocess the generation conditions to generate a target feature vector, and identify each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library. The copyright risk control instruction generation module is used to encode each of the copyright risk factors to obtain a copyright risk vector, determine the comprehensive risk score corresponding to the copyright risk vector, and generate a copyright risk control instruction based on the comprehensive risk score and the copyright risk vector. The candidate generation result determination module is used to control the generation process according to the copyright risk control instructions during the generative content reasoning process in order to obtain the corresponding candidate generation results; the control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints; The target generation result determination module is used to compare the similarity of the candidate generation results based on a preset protected content library and a preset copyright risk feature library, determine the target risk score according to the corresponding comparison results, determine the target generation result that meets the preset copyright risk threshold based on the target risk score, and output the corresponding copyright risk control record.
[0013] Thirdly, this application discloses an electronic device, including: Memory, used to store computer programs; A processor for executing computer programs to implement the steps of copyright risk control methods in the generative content reasoning process described above.
[0014] Fourthly, this application discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the copyright risk control method in the aforementioned generative content reasoning process.
[0015] This invention first obtains generation conditions, preprocesses these conditions to generate a target feature vector, identifies copyright risk factors in the target feature vector based on a preset protected content library and a preset copyright risk feature library, encodes each copyright risk factor to obtain a copyright risk vector, determines the comprehensive risk score corresponding to the copyright risk vector, and generates a copyright risk control instruction based on the comprehensive risk score and the copyright risk vector. During generative content reasoning, the generation process is controlled according to the copyright risk control instruction to obtain corresponding candidate generation results. The control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints. The candidate generation results are compared for similarity based on the preset protected content library and the preset copyright risk feature library. A target risk score is determined based on the comparison results. A target generation result that meets a preset copyright risk threshold is determined based on the target risk score, and the corresponding copyright risk control record is output. As can be seen, this application, by extracting copyright risk factors from the generation conditions and encoding them into a unified copyright risk vector, can quantify and fuse explicit and implicit copyright risks scattered across different modal inputs, enabling the system to obtain a structured comprehensive risk measurement before inference. Based on this, it generates appropriate copyright risk control instructions based on the comprehensive risk score and the copyright risk vector. During the inference process, it directly intervenes in the generation process through various control operations such as adjusting the prompt word vector representation, attention weights, sampling trajectories, and negative constraints. This ensures that risk avoidance behavior occurs within content generation rather than solely relying on input filtering or post-output screening, thereby reducing the probability of generated results being highly similar to protected expressions at the source and during the process. Simultaneously, by comparing the similarity of candidate generated results and selecting those meeting the preset copyright risk threshold based on the target risk score, coupled with the output of copyright risk control records, it can effectively reduce the retry computation overhead caused by high-risk outputs and provide traceable risk control evidence while maintaining the usability of the generated content. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0017] Figure 1 This application discloses a flowchart of a copyright risk control method in the generative content reasoning process. Figure 2 This is a schematic diagram of a copyright risk control device in a generative content reasoning process disclosed in this application. Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Since input-side rule review primarily relies on keywords or entity libraries, rule review may fail to detect implicit descriptions provided by users using combinations of "color + clothing + props + composition + camera movement." Simply using a word list or blacklist cannot solve the implicit copyright anchoring problem. Furthermore, since prompt rewriting schemes typically only process text, while generative content tasks may simultaneously include reference images, reference videos, style words, music clips, and generation parameters, rewriting only text prompts may retain high-risk protected expressions in the reference materials or audio clips, necessitating a unified modeling of multimodal inputs. Finally, since post-generation similarity filtering is performed only after the candidate generation results are completed... Therefore, if candidate generated results repeatedly approach protected expressions, the system needs to repeatedly generate and discard results, leading to increased inference time, memory usage, and calling costs. It is necessary to change the sampling trajectory in advance during the generation process, rather than filtering only at the endpoint. Since the controllable parameters differ for different generation models, task types, and risk factors—for example, image diffusion models can adjust guiding weights, negative constraints, and sampling steps; video generation models also need to consider inter-frame consistency; and audio generation models need to consider rhythm and melody structure—fixed thresholds or fixed rewriting strategies are difficult to adapt to multiple scenarios. A dynamic mapping from copyright risk to inference control parameters needs to be established. To address these technical problems, this application discloses a copyright risk control method, apparatus, device, and medium in the generative content inference process. Before and during generative content inference, it can uniformly represent copyright risk factors scattered in text, images, videos, audio, and generation parameters as a computable copyright risk vector. Based on this copyright risk vector, the generation process is dynamically adjusted, ensuring that the generated results actively avoid high-risk protected expressions while maintaining the user's creative intent, reducing the computational cost of repeated retries after generation and the probability of highly similar outputs.
[0020] To facilitate understanding of the embodiments of this application, some terms involved in the embodiments of this application will be explained first. Generation conditions refer to the information input by the user or upstream system that drives the output content of the generation model, including prompts, reference images, reference videos, style words, character descriptions, shot descriptions, text semantics, music clips, audio rhythm, seed values, generation parameters, etc.; copyright risk factors refer to factors extracted from generation conditions or intermediate inference states that may induce the generation of protected expressions, including protected entities, style, composition, shot, text semantics, audio rhythm, melody clips, character appearance, logo graphics, scene combinations, etc.; copyright risk vector refers to the vector representation of multiple copyright risk factors encoded into a unified risk space, used to describe the intensity, direction, and combination relationship of the generation task on different risk dimensions; protected entities refer to entities that may correspond to existing copyrighted works, character images, brand logos, portraits, film and television scenes, works of art, musical works, literary works, or combinations thereof; risk dimensions refer to the component categories in the copyright risk vector, such as entity dimension, style dimension, composition dimension, shot dimension, text semantic dimension, audio rhythm dimension, melody dimension, reference material similarity dimension, etc.; intermediate inference states refer to the intermediate variables of the generation model during the inference process, including latent variables of the diffusion model, noise prediction, etc. Attention maps, video frame latent variables, audio spectrum representation, pre-decoding features, candidate frame sequences, etc.; Dynamic avoidance refers to adjusting the sampling trajectory, prompt word rewriting results, guiding weights, negative constraints, and candidate generation result selection strategies in real time or in stages based on the copyright risk vector and the intermediate state of inference, so that the output content deviates from high-risk protected expressions; safe prompt words refer to prompt words obtained by de-entityizing, abstracting, or replacing the original prompt words, which retain the unprotected semantics in the user's creative intent while reducing the degree of anchoring to protected expressions; negative constraints are used to suppress specific... Constraints arising from entities, visual features, shot combinations, audio features, or textual expressions can be manifested as negative prompts, negative embeddings, loss terms, sampling penalty terms, or posterior filtering conditions. Copyright risk control instructions refer to the set of model control parameters derived from copyright risk vectors, including sampling weights, guiding weights, negative constraint weights, rewrite strength, frame-level avoidance strategies, and audio rhythm perturbation strategies. Similarity verification refers to comparing the similarity of candidate generated results with a preset protected content library or a preset copyright risk feature library during or after generation to confirm whether dynamic avoidance has reached a preset risk threshold. Furthermore, the risk library in this application refers to a database set used for copyright risk identification, similarity verification, control parameter generation, and verification result write-back. The risk library includes at least a preset protected content library and a preset copyright risk feature library, and may also include an entity alias library, style feature library, composition template library, shot template library, audio fingerprint library, and combined risk rule library.
[0021] See Figure 1As shown in the figure, this invention discloses a method for controlling copyright risks in generative content reasoning, including: Step S11: Obtain the generation conditions, preprocess the generation conditions to generate a target feature vector, and identify each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library.
[0022] In this embodiment, the generation conditions are first obtained. These conditions are information input by the user or upstream business system to drive the output content of the generation model. Specifically, they may include any one or a combination of several of the following: prompt words, reference images, reference videos, style words, text semantics, shot descriptions, music clips, audio rhythm, seed value, resolution, frame rate, generation duration, sampling steps, guidance strength, model type, and risk control strength selected by the user. It is understood that when obtaining the generation conditions, the system can assign a source identifier, timestamp, modality type, user configuration, and task identifier to each input, and record whether the input source is a direct user upload, a platform template, a reused historical project, or a third-party interface call, for use later when determining source reliability. For example, the system can provide the following input interfaces: prompt_text represents the text prompt word entered by the user; reference_images represents the set of reference images uploaded or selected by the user; reference_videos represents the set of reference videos uploaded or selected by the user; reference_audio represents the music clip or audio condition uploaded or selected by the user; style_terms represents the style words entered or selected by the user; generation_params represents generation parameters such as model type, resolution, frame rate, duration, number of sampling steps, and random seed; and risk_control_level represents the risk control strength configured by the user, platform, or enterprise.
[0023] After obtaining the generation conditions, multimodal preprocessing is performed on them, namely, standardizing the text, image, video, audio, and parameters respectively to obtain processed data such as vector representations of text semantics, image visuals, video frame sequences, audio spectrum, rhythm features, melody contour features, and generation parameter features. The processed data is then concatenated or fused to obtain the target feature vector. If the processed data contains only one vector representation, that vector representation is directly determined as the target feature vector. It should be noted that the target feature vector can be a single concatenated vector or a structured feature set composed of multiple modal sub-vectors and their index relationships; this application does not limit its specific organizational form.
[0024] In this embodiment, the preset protected content library is used to store content or content summaries that can be used as risk references. This may include, but is not limited to: work titles, character names, brand logos, character portraits, image works, video clips, music clips, lyric clips, literary text clips, posters, scene screenshots, storyboard clips, style samples, and manually labeled risk tags. The preset copyright risk feature library is used to store machine-processable features calculated from the preset protected content library, including vector representations of text semantics, lists of entity aliases, global vector representations of images, local visual fingerprints, contour descriptions, color histograms, protected expression diagrams, composition relationship diagrams, video shot sequence features, motion trajectory features, audio fingerprints, rhythm features, melody contours, chord progressions, and style tags. That is, the preset protected content library stores original or summarized objects, while the preset copyright risk feature library stores feature objects that can be used for rapid recall and similarity calculation; the two are associated through a protected content identifier (protected_reference_id). Furthermore, to reduce false positives, the preset copyright risk feature library can distinguish between and store ordinary public expressions and identifiable protected expressions. For example, "street in the rain," "neon lights," and "looking back" can be considered common public expressions; while "the composition path, character positioning, color scheme, and camera movement combination of consecutive shots in a specific film" can be considered high-risk combinations. When calculating the copyright risk vector, the system prioritizes assigning higher context enhancement coefficients to high-risk combinations, thus focusing on truly identifiable protected expressions without inadvertently harming common creative expressions.
[0025] After obtaining the target feature vector, the system identifies various copyright risk factors in the target feature vector based on a preset protected content library and a preset copyright risk feature library. Specifically, the system can extract copyright risk factors according to different modalities and risk types: For text prompts, the system performs word segmentation, entity recognition, alias matching, semantic vector recall, and large-scale model semantic verification. Entity recognition is used to discover explicit risks such as directly appearing work names, character names, brand names, and person names; alias matching is used to discover indirect entities such as nicknames, transliterations, abbreviations, and combined descriptions; and semantic vector recall is used to discover implicit anchors where "the entity name does not appear but the description is highly similar to the entity features." For reference images, the system extracts the main object, color combination, clothing and props, logo graphics, compositional relationships, visual style, edge contours, local visual fingerprints, and global visual embeddings. If the reference image is similar to a work, character, or logo in the preset protected content library in multiple visual dimensions, a reference material similarity risk factor is generated. For reference videos, the system is split into keyframes, shot segments, and action segments, and extracts character appearance, camera movement, scene layout, and action trajectory. The system employs various methods to detect and manage video elements. For example, it considers frame composition variations and video semantic embedding. If several consecutive frames are similar to protected video clips in shot order, composition path, or action rhythm, a video shot risk factor is generated. For music clips or audio cues, the system extracts beat, tempo, accent distribution, melody outline, chord progression, timbre characteristics, and spectral fingerprints. If an audio clip is similar to a clip in the music library in rhythm, melody, or spectral characteristics, an audio rhythm risk factor or melody risk factor is generated. For style terms, the system distinguishes between general style descriptions and style expressions pointing to specific authors, works, or brand visual systems. If a style term is strongly correlated with a specific protected subject, it does not simply delete all styles but decomposes them into retrievable general visual attributes and protected expression attributes that need to be avoided. For generation parameters, the system determines generation parameter risk factors based on the association rules between generation parameters such as seed value, resolution, frame rate, generation duration, sampling steps, model type, and guidance strength, and the protected content.
[0026] The record information corresponding to each copyright risk factor may include: risk factor identifier (factor_id), generation task identifier (task_id), risk source modality (including any one of text, image, video, audio, and generation parameter), risk type (including any one of protected entity, style, composition, shot, text semantics, audio rhythm, melody, reference similarity, and generation parameter), risk source location (e.g., character range of prompt words, coordinate frame of reference image, video frame range, audio time range), normalized risk expression (normalized_expression), associated protected content identifier (protected_reference_id), similarity value (similarity_value), confidence value (confidence_value), risk severity level (severity_level), and suggested action (suggested_action). For example, for the text "red cheeks, lightning tail, yellow electric animal", the system can generate three basic risk factors and one combined risk factor. The basic risk factors correspond to color appearance, tail shape and entity semantics, respectively. The combined risk factor records the common occurrence of these basic factors in the same subject description, so its context enhancement coefficient is higher than any single factor.
[0027] Step S12: Encode each of the copyright risk factors to obtain a copyright risk vector, determine the comprehensive risk score corresponding to the copyright risk vector, and generate a copyright risk control instruction based on the comprehensive risk score and the copyright risk vector.
[0028] In this embodiment, the system maps each copyright risk factor to a unified risk space and encodes it into a copyright risk vector R: ; Where, r e Indicates the risk component of the protected entity, r s Indicates the style risk component, r c Represents the compositional risk component, r l Indicates the risk component of the lens, r t r represents the semantic risk component of the text. a Indicates the audio rhythm risk component, r m Indicates the melody risk component, r refIndicates the risk component similar to the reference material, r p This represents the risk components of the generated parameters. Each risk component can be calculated using the following formula: ; Where i represents the risk dimension; w i This indicates the preset or adaptive weights for that risk dimension; sim i This indicates the similarity (i.e., second similarity) between the generated conditions and the preset protected content library in this dimension; conf i Indicates the confidence level of risk identification; src i This indicates the reliability of the source, which can be direct user input, parsed reference materials, semantic recall, or intermediate inference states; ctx i This represents the context enhancement coefficient, used to indicate the cumulative risk when multiple risk factors are combined. The overall risk score is... R It can be represented as: ; Where σ represents the normalization function, β i γ represents the first-order weight of each risk dimension. i,j This represents the interaction weight between two risk dimensions i and j. By incorporating interaction terms, the system can identify scenarios where "a single factor is not high-risk, but multiple factors combined form a high-risk protected expression." For example, yellow, round face, electric type, and animated animal may not be high-risk on their own, but when these factors are combined with specific tail shapes, cheek colors, and fighting postures, the risk score increases.
[0029] The data structure for a copyright risk vector can include fields as shown in Table 1. This data structure can be stored as vectors and metadata in a vector database, or as a relational table, key-value table, or message queue event. Its purpose is to transform copyright risk from a single threshold into a structured object that includes the source of risk, risk dimensions, confidence level, and actionable avoidance measures.
[0030] Table 1. Data Structure Diagram of Copyright Risk Vector
[0031] After obtaining the comprehensive risk score, a copyright risk control instruction C is generated based on the comprehensive risk score and the copyright risk vector. The copyright risk vector is mapped to a mixture of weights α (including vector representations of the original prompt words and vector representations of the security prompt words), a generative model guidance strength λ, a negative constraint weight μ, a sampling trajectory offset strength ν, and a security prompt word P. safe Negative constraint set P neg Similarity verification threshold T review and the number of local resampling times K resampleat least two of the copyright risk control instructions; wherein, if the comprehensive risk score is less than a first preset threshold T1, the original prompt and reference material are retained, and only light-weight review is performed; if the comprehensive risk score is greater than or equal to T1 but less than a second preset threshold T2, a safe prompt is generated and the weight of the negative constraint is increased; if the comprehensive risk score is greater than or equal to T2, the influence of the vector representation corresponding to the high-risk segment in the original prompt is reduced, the influence of the vector representation of the safe prompt and the weight of the negative constraint are increased, and intermediate reasoning state review is performed. The copyright risk control instruction C can be expressed as: ; wherein α represents the mixing weight of the vector representation of the original prompt (i.e., original prompt embedding) and the vector representation of the safe prompt (i.e., safe prompt embedding); λ represents the generation model guidance strength; μ represents the negative constraint weight; ν represents the sampling trajectory offset strength; P safe represents the safe prompt; P neg represents the negative constraint set; T review represents the similarity review threshold; K resample represents the number of local resamplings or the number of candidates. That is to say, the system converts the copyright risk vector into executable control parameters such as prompt rewriting strength, original prompt embedding weight, safe prompt embedding weight, negative constraint weight, sampling step adjustment strategy, candidate frame screening strategy, audio rhythm perturbation strategy, and post-generation review threshold.
[0032] Further, the system may set three risk thresholds T1, T2 and T3 to form a hierarchical control strategy: when Score R <T1, the task is a low-risk task, the system retains the original generation conditions, only records the risk control log and performs light-weight post-generation review; when T1≤Score R <T2, the task is a medium-risk task, the system performs safe prompt generation, negative constraint generation and post-generation review; when T2≤Score R <T3, the task is a high-risk task, the system enables intermediate reasoning state review, dynamically adjusts the sampling trajectory and performs local resampling; when Score R ≥T3, the task is an extremely high-risk task, the system may refuse generation, require the user to modify the input, or only allow generation of abstract, non-figurative results. The above thresholds can be flexibly configured according to the model type, business scenario, user permission, content type and regional policy. For example, a corporate internal material library may adopt a lower threshold to avoid generating similar commercial materials; a personal draft scenario may adopt a higher threshold and rely more on prompt modification; a public video release scenario may adopt the strictest threshold.
[0033] Step S13: During the generative content reasoning process, the generation process is controlled according to the copyright risk control instructions to obtain the corresponding candidate generation results; the control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints.
[0034] In this embodiment, during diffusion sampling, video frame generation, audio segment generation, or multimodal decoding, the sampling trajectory is adjusted according to copyright risk control instructions, and the control strength is dynamically updated based on periodic risk changes during inference intermediate states. Taking an image diffusion model as an example, the system can use the following control methods to achieve dynamic avoidance at the sampling trajectory level: ; Where, ε control This represents the noise prediction after copyright risk control; ε base Indicates the basic noise prediction; ε safe Noise prediction under the condition of safety warning words; ε orig This represents noise prediction under the original cue word conditions; ε uncond Indicates unconditional noise prediction; ε neg ΔR represents the noise direction corresponding to the negative constraint; ΔR represents the sampling trajectory offset direction determined by the copyright risk vector. Unlike simply adjusting the prompt words, this application transforms the copyright risk vector into a control term at the sampling trajectory level, causing the avoidance action to occur within the generation process. In this way, even if the original prompt words still retain the theme or emotion of the user's intent, the model will avoid high-risk entities, compositions, or style combinations during the latent variable evolution process. Furthermore, for entity words, style words, or composition words marked as high-risk, the system can also reduce their attention weight in cross-attention, i.e., perform attention weight adjustment, weakening the anchoring effect of high-risk segments on the generated results.
[0035] Step S14: Compare the similarity of the candidate generated results based on the preset protected content library and the preset copyright risk feature library, determine the target risk score according to the corresponding comparison results, determine the target generated result that meets the preset copyright risk threshold based on the target risk score, and output the corresponding copyright risk control record.
[0036] In this embodiment, similarity verification is not the sole control measure, but rather a verification step in the dynamic avoidance closed loop. The system performs multi-dimensional similarity comparisons between the candidate generated results and the preset protected content library and preset copyright risk feature library to obtain a target risk score (i.e., the verification risk score). If the target risk score exceeds the preset copyright risk threshold, high-risk areas are located based on the comparison results, and incremental avoidance operations such as local redrawing, segment resampling, or section replacement are performed on the located local areas. At the same time, the location results are written back to the copyright risk vector, increasing the corresponding risk component and driving the next round of control. If the target risk score is lower than the preset copyright risk threshold, the candidate generated result is determined as the target generated result and output. Simultaneously, the system outputs corresponding copyright risk control records. These records may include: the standardized original prompt (normalized_prompt), the safe prompt (safe_prompt), the negative constraint set (negative_constraints), the copyright risk vector (risk_vector), the copyright risk control instruction (control_instruction), the intermediate review result of the inference intermediate state, the review risk score of the candidate generated result (final_review_score), the retained user creative intent summary (retained_intent), and the modified, weakened, or avoided risk fragments (modified_fragments). These copyright risk control records can be used for auditing, model debugging, enterprise content security compliance, manual review, and risk database updates, making the entire avoidance process explainable and traceable.
[0037] In practical engineering implementation, the above method can be implemented using a four-layer technical architecture. The first layer is the generation condition access layer, which receives multimodal generation conditions submitted by users or upstream business systems and assigns a source identifier, timestamp, modality type, user configuration, and task identifier to each input. The second layer is the risk modeling layer, which constructs a set of copyright risk factors and a copyright risk vector. This layer includes a preset protected content library, a preset copyright risk feature library, an entity alias library, a style feature library, a composition template library, a shot template library, an audio fingerprint library, and a combined risk rule library. The output of the risk modeling layer is not a single rejection result, but a structured risk object with source, dimension, confidence level, and suggested avoidance actions. The third layer is the inference control layer, which generates copyright risk control instructions based on the copyright risk vector and injects these instructions into the generation model's inference process. For white-box or controllable models, the inference control layer can directly adjust the prompt word embedding, attention weight, noise prediction, guidance weight, and sampling step strategy. For black-box models, the inference control layer can indirectly achieve control through safety prompt words, negative prompt words, reference material weight reduction, candidate number adjustment, and local retry strategies. The fourth layer is the verification and write-back layer, which is used to verify the similarity of intermediate inference states, low-resolution previews, candidate images, candidate video clips, or candidate audio clips, and write the verification results back to the copyright risk vector or risk library, so that the system has closed-loop capability, rather than one-time input filtering.
[0038] From the perspective of data flow between steps: Step S11 takes prompt words, reference images, reference videos, music clips, style words, and generation parameters as input, and obtains the generation task object through generation task parsing, modality recognition, and parameter validity checks. Then, through text segmentation, visual encoding, video slicing, audio feature extraction, and parameter standardization, a multimodal feature set (i.e., target feature vector) is obtained. Finally, through entity matching, vector recall, structure matching, and combination rule calculation, the copyright risk factor set is output. Step S12 takes the copyright risk factor set as input, and through risk dimension normalization, confidence fusion, context enhancement, and vector grouping... The process involves obtaining a copyright risk vector and a comprehensive risk score, and then generating a copyright risk control instruction through control strategy selection, parameter mapping, security prompt word generation, and negative constraint generation. Step S13 takes the copyright risk control instruction, the generation model, and the intermediate inference state as input, and outputs candidate generation results or intermediate verification events through sampling trajectory, attention weight, guidance weight adjustment, and local resampling. Step S14 takes the candidate generation results or intermediate verification events as input, and outputs the target risk score and write-back event through multi-dimensional similarity verification, risk localization, and threshold judgment, and finally outputs the target generation result and copyright risk control record. In summary, this embodiment achieves proactive avoidance of high-risk protected expressions while maintaining the user's creative intent through a continuous technical chain of risk factor extraction, risk vector construction, control instruction mapping, dynamic avoidance in inference, and verification and write-back, while reducing the computational cost of repeated retries after generation and the probability of high similarity output.
[0039] As can be seen, this application, by extracting copyright risk factors from the generation conditions and encoding them into a unified copyright risk vector, can quantify and fuse explicit and implicit copyright risks scattered across different modal inputs, enabling the system to obtain a structured comprehensive risk measurement before inference. Based on this, it generates appropriate copyright risk control instructions based on the comprehensive risk score and the copyright risk vector. During the inference process, it directly intervenes in the generation process through various control operations such as adjusting the prompt word vector representation, attention weights, sampling trajectories, and negative constraints. This ensures that risk avoidance behavior occurs within content generation rather than solely relying on input filtering or post-output screening, thereby reducing the probability of generated results being highly similar to protected expressions at the source and during the process. Simultaneously, by comparing the similarity of candidate generated results and selecting those meeting the preset copyright risk threshold based on the target risk score, coupled with the output of copyright risk control records, it can effectively reduce the retry computation overhead caused by high-risk outputs and provide traceable risk control evidence while maintaining the usability of the generated content.
[0040] Based on the previous embodiment, this application further explains the preprocessing of the generation conditions and the identification process of copyright risk factors.
[0041] In one specific implementation, the preprocessing of the generation conditions to generate the target feature vector may include: standardizing the text, image, video, audio, and parameters respectively to obtain processed data, and then concatenating or fusing the processed data to obtain the target feature vector; wherein, if the processed data contains only one vector representation, then the vector representation is directly determined as the target feature vector. For text, standardization processing can include case normalization, full-width / half-width normalization, word segmentation, stop word removal, and calculating the semantic vector representation of the text using a text encoder. For images, standardization processing can include size normalization, color space normalization, and calculating the visual vector representation of the image using a visual encoder. For videos, standardization processing can include frame extraction, slicing by keyframes, shot segments, and action segments, and calculating the vector representation of the video frame sequence. For audio, standardization processing can include resampling, loudness normalization, and obtaining the vector representation of the audio spectrum using short-time Fourier transform, and further extracting rhythmic features such as beat, tempo, and accent distribution, as well as melodic contour features. For generation parameters, standardization processing can include encoding discrete or continuous parameters such as seed value, resolution, frame rate, generation duration, sampling steps, guidance strength, and model type into generation parameter features.
[0042] Regarding the specific methods of splicing or fusion, this application provides several optional implementations: In the first implementation, the system directly splices the vector representations of each modality in a preset order and adds modality identifier bits to obtain a single target feature vector. This method is simple to implement and requires no additional training. In the second implementation, the system performs weighted fusion of each modality vector according to modality importance. For example, for image-to-image tasks with a reference image as the main element, the fusion weight of the image visual vector representation is increased. For example, for the image generation task of "text prompt word + reference image", the system can splice a 768-dimensional text semantic vector and a 1024-dimensional image visual vector into a 1792-dimensional target feature vector, and record the modality and source location corresponding to each interval in the metadata.
[0043] As can be seen, because this embodiment standardizes the generation conditions of different modalities and dimensions into a computable vector representation and organizes them into a target feature vector, subsequent risk identification can be performed within a unified feature space, without having to design isolated review rules for each modality. Furthermore, since the target feature vector retains the correspondence between each modality's sub-features and their source locations, the subsequently identified risk factors can be accurately traced back to prompt word fragments, reference image regions, video frames, or audio segments, providing a data foundation for fine-grained avoidance control. Compared to existing input-side rule review methods that only perform word-based matching on text, this embodiment fundamentally expands the input surface for risk identification and improves its adaptability to multimodal generation tasks.
[0044] In one specific implementation, the step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library may include: determining the feature to be identified corresponding to at least one modality among text, image, video, and audio from the target feature vector; performing at least one of entity matching, alias matching, vector recall, structural matching, and combination rule calculation on the feature to be identified based on machine-processable features in the preset copyright risk feature library to determine the protected content object matched by the feature to be identified, and a first similarity between the feature to be identified and the matched protected content object; and generating each copyright risk factor of the matched protected content object based on the first similarity and the risk level, work type, and ownership information of the matched protected content object recorded in the preset copyright risk feature library.
[0045] Among them, entity matching refers to matching the entity words in the feature to be identified with the work name, character name, brand name, and person name in the preset copyright risk feature library to discover explicit risks; alias matching refers to identifying indirect entities such as nicknames, transliterations, abbreviations, foreign spellings, or combined descriptions based on the entity alias list; vector recall refers to performing nearest neighbor search on the vector representation of the feature to be identified in the preset copyright risk feature library to discover implicit anchors that "do not appear in the entity name but the description is highly close to the entity feature", and the first similarity can be the cosine similarity of the recall results; structural matching refers to performing graph matching or sequence alignment of structured features such as composition relationship graphs, video shot sequence features, motion trajectory features, or melody contours with protected structures in the library to discover structural risks at the composition, shot, and melody levels; combination rule calculation refers to judging whether multiple low-risk features appear together in a specific combination relationship according to the pre-configured combination risk rule library to discover combination risks.
[0046] It should be noted that the extraction of risk factors includes, but is not limited to, the identification methods described above, and this application does not limit them.
[0047] As can be seen, this embodiment comprehensively utilizes multiple identification methods such as entity matching, alias matching, vector recall, structural matching, and combination rule calculation, and organizes the identification results into structured risk factors with source location, confidence level, severity level, and suggested avoidance actions. Therefore, compared with conventional entity identification and blacklist filtering such as "identifying the work name or character name in the prompt words" in the prior art, this embodiment can hit both explicit entities and aliases, implicit semantics, and structural expressions. Furthermore, since the risk factors carry risk level, work type, and rights ownership information, the subsequent control strength can match the actual protection status of the hit object, avoiding unnecessary avoidance of authorized materials or low-risk public expressions, thereby improving risk coverage while reducing false alarm rate.
[0048] Furthermore, this embodiment of the invention also discloses a process for identifying implicit anchoring risks. Implicit anchoring risks refer to situations where the generation conditions do not directly include the work name, character name, brand name, or author name, but point to a protected expression through a combination of multiple common descriptions. Specifically, the step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library may further include: using at least two of the following attributes corresponding to the target feature vector—color, shape, clothing, props, actions, scenes, shots, timbre, and rhythm—as nodes, and using at least one of the same subject, the same shot, consecutive frames, and the same audio window as edges between the nodes to construct an attribute combination graph; performing subgraph matching or graph vector representation similarity comparison between the attribute combination graph and the protected expression graph in the preset copyright risk feature library; generating a combination risk factor when the similarity between multiple nodes in the attribute combination graph and the protected expression graph in terms of combination relationships meets preset conditions; and determining the context enhancement coefficient of the combination risk factor based on at least one of the following: the number of hit nodes, node rarity, relationship matching degree, and source aggregation degree.
[0049] Furthermore, since the same protected expression may be triggered simultaneously by text, images, videos, and audio, the system also performs multimodal risk alignment: copyright risk factors identified from different modalities are mapped to the same protected content identifier. Then, the cross-modal hit rate under the same protected content identifier is statistically analyzed. Cross-modal consistency is determined based on the ratio of the number of modalities hitting the same protected content identifier to the total number of modalities involved in the generation conditions. The context enhancement coefficient of the corresponding copyright risk factor is adjusted based on this cross-modal consistency. Cross-modal consistency can be expressed as: ; Among them, cross j Represents the cross-modal consistency of the j-th protected reference object, n j This indicates the number of modes that hit the protected reference object, where N represents the total number of modes involved in the current task, and cross. j The higher the value, the greater the context enhancement coefficient of the corresponding risk component. For example, if the user text prompt includes "classic space opera villain armor," the reference image shows a black helmet outline, and the reference video shows a similar cape turning shot, each item individually may be below the high-risk threshold. However, when all three refer to the same protected character or work, n... j =3, N=3, cross-modal consistency is 1, the system uses this as a risk enhancement signal to significantly increase the context enhancement coefficient, thus making the combination a high-risk combination.
[0050] As can be seen, because this embodiment identifies combination risks by matching attribute combination graphs with protected expression graphs and adjusts the context enhancement coefficients through cross-modal consistency, the system can detect implicit anchoring risks such as "individual descriptions are all common expressions, but when combined, they point to a specific protected expression," and cross-modal risks such as "low risk in a single modality, but multiple modalities pointing to the same object." Existing input rule-based review relies on explicit word lists, which are largely ineffective against these two types of risks. Existing prompt word rewriting schemes also fail to establish cross-modal associations because they only process text. Therefore, it can be deduced that by improving the granularity of risk identification from isolated words to combination and cross-modal relationships, the space for users to circumvent review through "color + clothing + props + composition + camera movement" is significantly reduced, and the recall rate of risk identification increases accordingly. Simultaneously, because common expressions have low context enhancement coefficients when there are no combination relationships, the false positive rate does not increase synchronously with the recall rate, achieving a balance between identification accuracy and coverage.
[0051] Based on the above embodiments, this invention further explains the process of constructing the copyright risk vector and generating the copyright risk control instructions.
[0052] In one specific implementation, encoding each copyright risk factor to obtain a copyright risk vector and determining the comprehensive risk score corresponding to the copyright risk vector may include: mapping each copyright risk factor to a unified risk space to obtain the copyright risk vector; and determining the comprehensive risk score corresponding to each risk component based on the weight w of the target risk dimension. i The second similarity sim between the target feature vector and the preset protected content library on the target risk dimension. i Confidence level identification i Source reliability src i and the context enhancement coefficient ctx i The product between them determines the risk component r. i Each risk component is assigned a first-order weight β based on the corresponding target risk dimension. i We perform weighting to obtain a first-order weighted result; the product r between any two different risk components is... i r j Based on the corresponding interaction weight γ i,j The weighted average is then applied to obtain the cross-weighted result; the sum of the first-order weighted result and the cross-weighted result is then normalized to obtain the comprehensive risk score. R .
[0053] It should be noted that the original scores for different risk dimensions have different dimensions. For example, visual similarity may come from cosine similarity, audio similarity from fingerprint matching score, text semantic similarity from vector recall score, and graph similarity from graph matching score. Therefore, before assembling the copyright risk vector, the system first normalizes the original scores for each dimension. Optional normalization implementations include: piecewise function mapping, max-min normalization, temperature scaling, and percentile mapping based on historical distribution. For example, if the visual similarity is higher than the 95th percentile in historical samples, it is mapped to a high-risk interval; if the audio rhythm similarity is high but the melody similarity is low, it is mapped using the dimension interaction weight γ. i,j Reduce overall audio copyright risks. Regarding the reliability of the source (src / source code)... i The value is determined based on the risk source modality and acquisition method of the copyright risk factor. For example, the value is 1.0 when the source is the text entity directly input by the user, 0.9 when the source is the parsing of reference materials, 0.7 when the source is the semantic vector recall, and 0.6 when the source is the inference of the intermediate state of reasoning.
[0054] When multiple risk factors exist within the same risk dimension, the system can calculate the components of that dimension using the maximum value, weighted average, or Top-K aggregation. Preferably, the system uses Top-K weighted aggregation: ; Where, q i,k ρ represents the normalized risk value of the k-th risk factor within the i-th risk dimension. k This indicates the ranking decay weight. This method can retain the impact of multiple risk factors occurring together, while avoiding false alarms caused by the simple accumulation of a large number of low-confidence factors.
[0055] As can be seen, since each risk component in this embodiment is obtained by multiplying five factors: dimension weight, second similarity, identification confidence, source reliability, and context enhancement coefficient, a low value of any factor will lower the component. Therefore, a single unreliable signal (such as semantic recall with low confidence) will not directly trigger high-intensity avoidance, making risk quantification more robust.
[0056] Furthermore, copyright risk control instructions can be generated according to different risk dimensions: For protected entity risks, the main risk is that the character, brand, person, work name or alias is directly or implicitly hit, and the corresponding control instructions are to reduce the entity word embedding weight, generate a generalized entity safety prompt word, and add negative entity constraints; For style risks, the main risk is that it points to a specific author, work or brand visual system, and the corresponding control instructions are to decompose the specific style into general style attributes and reduce the anchoring weight of the specific style; For composition risks, the main risk is that the subject position, color block ratio, prop combination, and layout relationship are close to the protected expression, and the corresponding control instructions are to adjust the composition description, change the subject ratio, perspective, prop position or background structure; For camera risks, the main risk is that the video shot order, camera movement, and character positioning are close to the protected expression. For protected segments, the corresponding control instructions are to change the camera angle, motion path, shot size, switch rhythm, or keyframe layout; for text semantic risks, the main risk is that the text description is similar to the protected storyline, dialogue, or setting, and the corresponding control instructions are to rewrite the plot, replace the setting, or delete identifiable dialogue structure; for audio rhythm risks, the main risk is that the beat, accent, or rhythm pattern is similar to a musical work, and the corresponding control instructions are to adjust the speed, accent, rhythm pattern, or measure structure; for melody risks, the main risk is that the melody outline or chord progression is similar to a musical work, and the corresponding control instructions are to perform melody shift, chord replacement, or interval perturbation; for reference material similarity risks, the main risk is that the uploaded reference image, video, or audio is similar to the protected content, and the corresponding control instructions are to reduce the weight of the reference material, crop high-risk areas, and abstract the reference conditions.
[0057] The control instruction generation module operates in the following order: first, it determines the risk level; then, it determines the creative intent to be preserved; next, it determines the expressive segments to be avoided; then, it selects at least one control action for each risk dimension; and finally, it merges multiple control actions into a single-inference executable control instruction. For cases where multiple risk dimensions are simultaneously affected, the system can generate a local control instruction C for each risk dimension. i Then, multiple local control instructions are merged into a global control instruction: ; The Merge function is sorted according to risk level, confidence level, user intent retention, and model controllability. All are local control commands. If multiple control actions conflict, such as one action requiring the preservation of the main character while another requires the deletion of the main character, or two control commands acting on the same prompt word fragment or the same visual area, the control command with the higher risk score and higher source reliability will be executed first, and the sacrificed user intent will be recorded in the copyright risk control record. If a control command will cause a significant decrease in the generated quality, the system can use lower-intensity control parameters and increase the number of post-generation reviews, using reviews as a fallback to maintain the generated quality.
[0058] In one specific implementation, generating the safety warning word may include the following sub-steps: First, the original warning word is semantically segmented to obtain multiple semantic segments such as theme segments, attribute segments, style segments, shot segments, entity segments, action segments, and purpose segments; Second, each semantic segment is aligned with copyright risk factors and marked as retainable segments, weakened segments, replaced segments, and deleted segments. Retainable intention segments include themes, scene categories, emotional atmospheres, common art styles, colors of non-specific objects, general camera language, and content purpose. Segments to be avoided include specific work titles, character names, specific brand logos, recognizable character appearance combinations, specific shot replicas, specific melody segments, and height-sensitive elements. The process involves six steps: 1) Associating the rewrite with a style from a specific author or work; 2) Abstracting the elements to be weakened to obtain abstract descriptions, such as abstracting a specific character's appearance as an "original fantasy creature," a specific author's style as "bright animation quality," and a specific film shot as "tense rainy night atmosphere"; 3) Generating alternative expressions for the elements to be replaced, such as changing color combinations, main outlines, prop shapes, camera angles, rhythms, or melodic directions; 4) Combining the reproducible elements, alternative expressions, and abstract descriptions into a safety prompt word and generating a set of negative constraints; 5) Calculating the intent retention score between the safety prompt word and the original prompt word. If the intent retention score is lower than a preset threshold (i.e., the third preset threshold), the rewrite intensity is reduced or the user is prompted for confirmation. Intent Retention Score S It can be represented as: ; Among them, P orig_nonrisk This refers to the non-risk creative intent segment after removing the expressions to be avoided (i.e., segments to be weakened, replaced, or deleted) from the original prompts. safe'sim' represents the semantic similarity calculation, and the formula ensures that the security prompt retains the user's context, purpose, mood, and general style requirements. It should be noted that the rewriting intensity is positively correlated with the difference between the security prompt and the original prompt; that is, the greater the rewriting intensity, the higher the degree of abstraction, replacement, and deletion of the security prompt relative to the original prompt, and the greater the semantic difference between the two. Therefore, reducing the rewriting intensity when the intent retention score is insufficient can directly reduce the deviation of the security prompt from the user's intent.
[0059] For example, the original prompt is "Generate a yellow electric creature resembling a classic cartoon character, with red cheeks and a lightning tail, standing in an arena". The system can identify the retainable intent as "cute small fantasy creature, electric theme, arena scene, bright animation texture" and the expression to be avoided as a combination of "yellow electric creature + red cheeks + lightning tail", and then generate the safety prompt "An original small fantasy creature with abstract electric elements and a non-marker tail outline, standing in a futuristic arena, with an overall bright animation texture, but without presenting the face, tail and color combination of a specific known character".
[0060] As can be seen, this embodiment explicitly splits the original prompt words into retainable intent fragments and avoidable expression fragments. The rewriting only applies to risky fragments. Since non-risk fragments are fully retained and quantitatively verified through intent retention scores, the generated result after rewriting still meets the user's theme, scenario, emotion, and usage needs, avoiding theme drift and usability reduction caused by overall rewriting.
[0061] Based on the above embodiments, this embodiment of the invention further describes the process of performing control operations to obtain candidate generation results during generative content reasoning. In this embodiment, the system does not calculate the risk only once before generation, but reviews the intermediate state of reasoning according to a preset step size or event triggering method during the reasoning process, and dynamically updates the copyright risk control instructions accordingly.
[0062] In one specific implementation, controlling the generation process according to the copyright risk control instruction to obtain corresponding candidate generation results may include: during generative content reasoning, extracting intermediate reasoning states based on a preset step size or event triggering method, wherein the intermediate reasoning states include at least one of intermediate latent variables, attention maps, low-resolution preview features of the current predicted image, video keyframe features, and audio segment features; comparing the similarity between the intermediate reasoning states and the preset copyright risk feature library to obtain a current intermediate risk score; updating the sampling trajectory offset intensity, the constraint intensity corresponding to the negative constraint, and the attention weight corresponding to the high-risk segment in the original prompt words according to the relationship between the current intermediate risk score and the previous intermediate risk score corresponding to the previous review step to obtain an updated copyright risk control instruction; and re-jumping to the step of extracting intermediate reasoning states according to the updated copyright risk control instruction until reasoning is completed to obtain corresponding candidate generation results.
[0063] For the diffusion model, the system can extract the latent variable x at the k-th sampling step. k Attention Map A k and the low-resolution preview features F of the current predicted image k , will F k A rapid similarity comparison is performed with a preset copyright risk feature database to obtain the current intermediate risk score M. k .
[0064] For dynamic updates of control parameters, if M k The intermediate risk score M is higher than that of the previous review step. k -1 and exceeds the intermediate threshold T mid This indicates that the sampling trajectory is approaching a high-risk protected expression, and the system increases the negative constraint weights and the sampling trajectory offset intensity. ; in, This indicates the negative constraint update step size. This indicates the sampling trajectory offset update step size. Simultaneously, the system reduces the attention weights corresponding to high-risk segments in the original prompts, weakening the anchoring effect of high-risk semantics on subsequent sampling. If M... k Below T mid As the number of attempts to avoid excessive avoidance leads to a decrease in the quality of the generated data, the system can gradually reduce the control intensity to prevent this. ; Where, η decay μ represents the decay step size. min and ν minThese represent the lower limits of the negative constraint weights and the sampling trajectory offset intensity, respectively. Through the above dynamic updates, the system can enhance avoidance when the risk increases and restore some user intent expression when the risk decreases, thereby achieving a balance between copyright risk control and generation quality.
[0065] For image generation scenarios, dynamic obstacle avoidance can specifically perform the following control operations: First, cue word embedding control, that is, embedding the original cue word into E orig Security prompts embedded in E safe and negative constraint embedding E neg The input diffusion model is adjusted based on the mixing weight α and the negative constraint weight μ to reduce the cross-attention weights of entity words, style words, or composition words marked as high-risk. Secondly, region-level avoidance is implemented: if a specific region in the reference image or intermediate preview image is similar to the protected content, a region mask is generated. risk In addition, negative constraint weights are increased or local redrawing is performed in the area, while the original generation conditions are maintained in non-risk areas to reduce overall image deviation; third, compositional perturbation, that is, if the compositional risk component is high, the position of the subject, perspective, shot size, relative position of props, color ratio or background elements are adjusted to make the image layout deviate from the protected expression, while retaining the theme and purpose; fourth, local resampling, that is, if only local areas of the final image are high risk, such as logos, facial features, tail shapes or prop outlines, local resampling is performed only on that area, instead of regenerating the entire image.
[0066] For video generation scenarios, in addition to handling single-frame visual similarity, the system also addresses shot and action risks over time. The system first breaks down the video generation task into keyframes, transition frames, and shot segments, and then establishes a segment-level risk vector R for each shot segment. seg If the shot risk component r of a certain shot segment l If the risk is high, the system can change the camera movement, such as changing a horizontal tracking shot to a fixed shot or a low-angle zoom. If the movement trajectory is high-risk, the system can change the start and end posture of the movement, the speed of movement, or the direction of turning. If the scene layout is high-risk, the system can change the background structure, the position of the light source, or the position of the subject. To maintain video continuity, the system applies consistency constraints to the boundaries of the local resampled segments. In other words, when avoiding high-risk shots, the system does not directly disrupt the subject identity, tone, and scene continuity in the preceding and following frames, but adjusts the specific shot structure of the high-risk protected expression within a controllable range.
[0067] For audio generation scenarios, the system divides the audio into multiple segments by measures, beat windows, or fixed time windows, and calculates the rhythmic and melodic risks of each segment. If the rhythmic risk is high but the melodic risk is low, the system prioritizes adjusting the accent distribution, drum density, or syncopation; if the melodic risk is high, the system prioritizes adjusting the interval direction, cadences, repetitive motifs, or chord progressions. For scenarios such as advertising background music and short video soundtracks, the system can retain tempo, mood, and timbre categories while changing the melody outline and rhythmic pattern. For scenarios where users upload reference audio, the system can first abstract the reference audio into retainable conditions such as "mood, tempo, instrument, and energy curve," and then use the melodic fingerprint and specific rhythmic patterns as expressions to be avoided.
[0068] It should be noted that the implementation of the above control operations depends on the controllability of the model. This application provides corresponding adaptation methods for white-box and black-box models. For white-box models or pluggable inference pipelines, the system can directly access the text encoder, attention layer, latent variables, sampler, and decoder. Therefore, copyright risk control instructions can be directly applied to cue word embedding, cross-attention weights, noise prediction direction, sampling step size, negative conditional embedding, and local redraw masking. For black-box models or third-party API (Application Programming Interface) models, the system cannot directly access the intermediate inference state. In this case, the copyright risk control instructions can be converted into externally controllable parameters, including safety cue words, negative cue words, reference material weights, candidate number, random seed perturbation, generation size adjustment, staged generation, and candidate generation result filtering. Even in black-box scenarios, this application can still achieve dynamic avoidance through pre-generation risk vector construction, input-side control instructions, and post-generation review and write-back. The only difference is that the intermediate inference state review is replaced by multi-candidate generation result review or low-resolution pre-generation review. In addition to the preferred inference control method of adjusting prompt word embedding, guiding weight, negative constraints and sampling trajectory, the generation process can also be controlled by alternative methods such as adjusting random seed, changing sampler, modifying noise schedule, adjusting attention weight, adjusting candidate number, setting local redrawing area or modifying decoding parameters. This application does not limit the scope of these methods.
[0069] Based on the above embodiments, the present invention further explains the process of comparing the similarity of candidate generation results and determining the target generation result.
[0070] In one specific implementation, the step of comparing the similarity of the candidate generated results based on a preset protected content library and a preset copyright risk feature library, determining a target risk score based on the corresponding comparison results, and determining a target generated result that meets a preset copyright risk threshold based on the target risk score may include: comparing the candidate generated results based on at least two of the following similarities: visual similarity, textual semantic similarity, audio similarity, structural similarity, and entity recognition similarity, to obtain corresponding comparison results; and weighting and summing the similarity scores in the comparison results based on corresponding preset weights to obtain the target risk score. The target risk score (i.e., the review risk score) Score0 can be represented as: ; Among them, S v S represents visual similarity. t S represents the semantic similarity of texts. a S represents audio similarity. str S represents structural similarity. ent Indicates entity recognition similarity; δ v δ t δ a δ str δ ent These represent the preset weights for each review dimension. For example, for image generation tasks, visual similarity, structural similarity, and entity recognition similarity can be mainly used; for video generation tasks, structural similarity at the shot sequence level can be added on top of the above; for audio generation tasks, audio similarity and structural similarity can be mainly used; and for text generation tasks, text semantic similarity and entity recognition similarity can be mainly used.
[0071] When Score0 exceeds the review threshold T review In such cases, the system does not need to completely regenerate; instead, it locates high-risk areas based on the review results. For example, for images, it can locate the main outline, logo, color scheme, or composition; for videos, it can locate frames, shots, or motion trajectories; and for audio, it can locate measures or melodic fragments. The system writes the location results back to the copyright risk vector, increasing the corresponding risk component and triggering local resampling or local replacement.
[0072] Furthermore, the review results are also written back to the copyright risk vector to form a closed-loop control. The system updates the corresponding risk components based on the target risk score and risk location results: ; Where, r i,new Let κ represent the updated risk component, and loc represent the write-back coefficient. iThis indicates the correlation between the reviewed location result and the i-th risk dimension. `clip` restricts the result to between 0 and 1. If a candidate generated result passes review, the system can appropriately reduce the corresponding risk component to avoid excessive avoidance during the next local resampling. Each write-back can be recorded as a write-back event, with fields including: review event identifier `review_id`, generation task identifier `task_id`, candidate generated result identifier `candidate_id`, risk location result `risk_location`, review risk score `review_score`, updated risk dimension `updated_dimension`, risk component update value `update_value`, and next action `next_action`. The next action includes passing, local resampling, overall resampling, rejecting generation, or transferring to manual review.
[0073] As can be seen, the review in this embodiment is multi-dimensionally weighted (visual, textual semantics, audio, structural, and entity recognition, at least two). Since different task types can be configured with different review dimension weights, a target risk score matching its risk characteristics can be given for multi-form generation results such as images, videos, audio, and text, avoiding the omission of structural or entity risks by a single similarity index. This embodiment writes the review results back to the copyright risk vector according to the degree of relevance. Since the corresponding risk component increases after writing back, the control intensity of the next round will be automatically increased. And through review, the component can be recalled to avoid excessive avoidance. Therefore, the system can gradually converge to the balance point of risk and quality in multi-round generation and batch generation scenarios, overcoming the defects of conventional retries that "start from zero every time and have no experience accumulation".
[0074] To facilitate understanding, the complete execution process of the above method will be explained below using three specific application scenarios.
[0075] Application Scenario 1: Image Generation Scenario. User input: "Create an advertisement image for a children's toy. The reference image is a yellow electric animal with red cheeks, a lightning bolt-shaped tail, and a style similar to classic Japanese animation, standing in the center of an arena." After the system executes step S11 to obtain the prompt words and reference image, the text side identifies risk factors such as "yellow electric animal, red cheeks, lightning bolt-shaped tail, classic Japanese animation, arena," while the image side identifies risk factors such as the main body outline, cheek color blocks, tail shape, and color combination. The system executes step S12 to calculate the risk component r of the protected entity. e =0.83, Style Risk Component r s =0.62, Graph risk component r c =0.35, Reference material similarity risk component r ref =0.88, Overall Risk Score R=0.79, which is higher than the second preset threshold T2=0.70, so the task is judged as a high-risk task. The system generates the following safety warning words: "Original small fantasy creature with abstract electrical energy elements, non-marker tail outline, bright animated advertising texture, standing in the center of a future arena, avoid presenting the face, tail and color combination of a specific known character", and generates a set of negative constraints: "specific red cheek combination, lightning tail, known yellow electric character outline, specific brand character proportion", etc. The system executes step S13 to perform reasoning. When the similarity between the main outline of the preview image in step 20 and the risk library increases, the system increases the negative constraint weight μ and adjusts the sampling trajectory offset intensity ν, while reducing the attention weight of high-risk segments such as "lightning tail". The system executes step S14 to perform visual, structural, and entity recognition similarity verification on the candidate images. If the target risk score is lower than the preset copyright risk threshold, the final output is an original toy advertisement image with an electric energy theme but different in outline, face proportion, tail shape, and color scheme. The system also outputs a copyright risk control record that records the rewritten fragment, negative constraint terms, sampling control parameters, and verification score.
[0076] Application Scenario 2: Video Generation. User input: "Generate a shot of a protagonist looking back with an umbrella in a rainy city night, referencing a neon street scene from a movie; the shot must closely resemble it." The system executes step S11 to extract camera movement, composition, color layout, and protagonist's movement trajectory from the reference video, identifying shot risk factors and composition risk factors. The system executes step S12, breaking down the risk expression of "closely resembling the reference video" before generation into retainable elements such as "rainy night, neon lights, looking back, and emotional tension," and elements to be avoided such elements such as "specific camera angles, street layouts, character positioning, and camera movement paths," and generates corresponding copyright risk control instructions. The system executes step S13 to perform video reasoning, reviewing intermediate states by keyframe. When it finds that the camera path of consecutive frames is close to the protected segment, it changes the camera height, focal length, and protagonist's movement direction, and applies consistency constraints to the front and back boundaries of the locally resampled segment, maintaining the rainy night mood while avoiding replicating specific shots. The system executes step S14 to verify the frame-level visual similarity and shot sequence structure similarity of the candidate video. After confirming that the target risk score is below the threshold, it outputs the target video and copyright risk control record.
[0077] Application Scenario 3: Music Generation. A user uploads a music clip and inputs: "Generate an advertising background music with the same rhythm and similar melody as this clip." The system executes step S11 to extract rhythmic features, melody outlines, and spectral fingerprints, identifying audio rhythm and melody risk factors. The system executes step S12, retaining the general creative intent of "advertising background music, upbeat, electronic timbre, approximately 120 BPM (Beat Per Minute)," and incorporates specific melodic direction and strongly similar rhythmic patterns into negative constraints, generating copyright risk control instructions. The system executes step S13 to perform audio reasoning, checking melody similarity by measure window. If the risk of a certain window increases, chord replacement, melody shift, and rhythmic perturbation are performed to ensure the generated audio maintains the upbeat advertising atmosphere but deviates from the original clip's expression. The system executes step S14 to check the audio similarity and structural similarity of candidate audio. If a measure still exceeds the threshold, only that audio window is replaced by measure and checked again. Finally, the system outputs the target audio that meets the preset copyright risk threshold and the copyright risk control record.
[0078] Furthermore, it should be noted that the method provided in this application can be deployed as a server-side plugin for a generation platform, middleware for model inference services, an enterprise content security gateway, a client-side SDK (Software Development Kit), a browser extension, or a cloud API service. That is, besides the preferred deployment on the inference service layer of the generation platform, it can also be deployed as a browser plugin, API gateway, enterprise content security proxy, client-side SDK, or cloud content compliance service. For different deployment forms, the core steps remain copyright risk factor extraction, copyright risk vector construction, copyright risk control instruction generation, dynamic avoidance inference, and similarity verification and write-back. Similarly, in addition to the preferred representation of a fixed-dimensional risk vector including confidence, source reliability, and context enhancement coefficients, the copyright risk vector can also adopt alternative representations such as risk graphs, risk matrices, sparse vectors, hierarchical label trees, probability distributions, or state machines. This application does not limit these representations.
[0079] See Figure 2 As shown, this embodiment of the invention discloses a copyright risk control device in the generative content reasoning process, comprising: The copyright risk factor identification module 11 is used to obtain the generation conditions, preprocess the generation conditions to generate a target feature vector, and identify each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library. The copyright risk control instruction generation module 12 is used to encode each of the copyright risk factors to obtain a copyright risk vector, determine the comprehensive risk score corresponding to the copyright risk vector, and generate a copyright risk control instruction based on the comprehensive risk score and the copyright risk vector. The candidate generation result determination module 13 is used to control the generation process according to the copyright risk control instruction during the generative content reasoning process to obtain the corresponding candidate generation result; the control operation includes any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectory, and enhancement of negative constraints. The target generation result determination module 14 is used to compare the similarity of the candidate generation results based on the preset protected content library and the preset copyright risk feature library, determine the target risk score according to the corresponding comparison results, determine the target generation result that meets the preset copyright risk threshold based on the target risk score, and output the corresponding copyright risk control record.
[0080] Since the embodiments of the device part correspond to the embodiments described above, please refer to the embodiments described in the method part for the embodiments of the device part, and will not be repeated here.
[0081] As can be seen, this application, by extracting copyright risk factors from the generation conditions and encoding them into a unified copyright risk vector, can quantify and fuse explicit and implicit copyright risks scattered across different modal inputs, enabling the system to obtain a structured comprehensive risk measurement before inference. Based on this, it generates appropriate copyright risk control instructions based on the comprehensive risk score and the copyright risk vector. During the inference process, it directly intervenes in the generation process through various control operations such as adjusting the prompt word vector representation, attention weights, sampling trajectories, and negative constraints. This ensures that risk avoidance behavior occurs within content generation rather than solely relying on input filtering or post-output screening, thereby reducing the probability of generated results being highly similar to protected expressions at the source and during the process. Simultaneously, by comparing the similarity of candidate generated results and selecting those meeting the preset copyright risk threshold based on the target risk score, coupled with the output of copyright risk control records, it can effectively reduce the retry computation overhead caused by high-risk outputs and provide traceable risk control evidence while maintaining the usability of the generated content.
[0082] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.
[0083] Figure 3This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the copyright risk control method during the generative content reasoning process disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.
[0084] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.
[0085] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0086] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the copyright risk control method in the generative content reasoning process executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.
[0087] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the copyright risk control method in the aforementioned generative content reasoning process. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.
[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.
[0089] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0090] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0091] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0092] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for controlling copyright risks in generative content reasoning, characterized in that, include: The generation conditions are obtained and preprocessed to generate a target feature vector. Based on a preset protected content library and a preset copyright risk feature library, each copyright risk factor in the target feature vector is identified. Each of the copyright risk factors is encoded to obtain a copyright risk vector, a comprehensive risk score corresponding to the copyright risk vector is determined, and a copyright risk control instruction is generated based on the comprehensive risk score and the copyright risk vector. During the generative content reasoning process, the generation process is controlled according to the copyright risk control instructions to obtain the corresponding candidate generation results; the control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints. The candidate generated results are compared for similarity based on a preset protected content library and a preset copyright risk feature library. A target risk score is determined based on the comparison results. Based on the target risk score, a target generated result that meets the preset copyright risk threshold is determined, and the corresponding copyright risk control record is output. The process of encoding each of the copyright risk factors to obtain a copyright risk vector and determining the comprehensive risk score corresponding to the copyright risk vector includes: The copyright risk factors are mapped to a unified risk space to obtain the copyright risk vector; the copyright risk vector includes at least two risk components from the following: protected entity risk component, style risk component, composition risk component, shot risk component, text semantic risk component, audio rhythm risk component, melody risk component, reference material similarity risk component, and generation parameter risk component. The risk component is determined by multiplying the weight of the target risk dimension corresponding to each risk component, the second similarity between the target feature vector and the preset protected content library on the target risk dimension, the identification confidence, the source reliability, and the context enhancement coefficient of the copyright risk factor; wherein, the source reliability is determined according to the risk source modality of the copyright risk factor, and when there are multiple copyright risk factors in the same risk dimension, the risk component corresponding to the risk dimension is determined according to the maximum value, weighted average, or Top-K weighted aggregation method; Each of the aforementioned risk components is weighted according to the first-order weight of the corresponding target risk dimension to obtain the first-order weighted result; The product of any two different risk components is weighted according to the corresponding interaction weights to obtain the interaction weighted result. The summation of the first-order weighted result and the interactive weighted result is normalized to obtain the comprehensive risk score.
2. The method of controlling copyright risk in a generative content reasoning process of claim 1, wherein, The generation conditions include any one or a combination of several of the following: prompt words, reference images, reference videos, style words, text semantics, shot descriptions, music clips, audio rhythm, seed value, resolution, frame rate, generation duration, number of sampling steps, guidance strength, model type, and risk control strength selected by the user. Accordingly, the preprocessing of the generation conditions to generate the target feature vector includes: The text, images, videos, audio, and parameters are standardized respectively to obtain processed data; the processed data includes any one or a combination of several of the following: vector representation of text semantics, vector representation of image vision, vector representation of video frame sequence, vector representation of audio spectrum, rhythm features, melody contour features, and generation parameter features. The processed data is concatenated or fused to obtain the target feature vector; wherein, if the processed data contains only one vector representation, then the vector representation is directly determined as the target feature vector.
3. The method of controlling copyright risk in a generative content reasoning process of claim 1, wherein, The copyright risk factors include any one or a combination of several of the following: protected entity risk factors, style risk factors, composition risk factors, shot risk factors, text semantic risk factors, audio rhythm risk factors, melody risk factors, reference material similarity risk factors, and generation parameter risk factors. The generation parameter risk factor is determined based on the association rule between at least one of the generation parameters in the generation conditions, namely seed value, resolution, frame rate, generation duration, sampling steps, model type, and guidance strength, and the protected content; the preset protected content library is a library that stores protected content objects; the preset copyright risk feature library is a library that stores machine-processable features calculated through the preset protected content library; The machine-processable features include any one or a combination of several of the following: vector representation of text semantics, entity alias list, global vector representation of image, local visual fingerprint, color histogram, protected representation diagram, composition relationship diagram, video shot sequence features, motion trajectory features, audio fingerprint, rhythm features, melody outline, and style tags; each machine-processable feature in the preset copyright risk feature library is associated with the corresponding protected content object in the preset protected content library through a protected content identifier; Accordingly, the step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library includes: From the target feature vector, determine the features to be identified corresponding to at least one modality among text, image, video, and audio; Based on the machine-processable features in the preset copyright risk feature library, at least one of entity matching, alias matching, vector recall, structure matching and combination rule calculation is performed on the feature to be identified to determine the protected content object hit by the feature to be identified, and the first similarity between the feature to be identified and the hit protected content object. Based on the first similarity and the risk level, work type and ownership information of the matched protected content object recorded in the preset copyright risk feature library, generate each copyright risk factor of the matched protected content object; The record information corresponding to each copyright risk factor includes any one or a combination of several of the following: risk factor identifier, corresponding generation task identifier, risk source modality, risk type, source location, standardized risk expression, associated protected content identifier, first similarity, identification confidence level, risk severity level, and suggested avoidance action; the risk source modality includes any one of text, image, video, audio, and parameters; the risk type includes any one of protected entity, style, composition, shot, text semantics, audio rhythm, melody, similarity of reference materials, and generation parameters.
4. The method of controlling copyright risk in a procedural content reasoning process of claim 3, wherein, The step of identifying each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library also includes: At least two of the following characteristics corresponding to the target feature vector—color, shape, clothing, props, actions, scenes, shots, timbre, and rhythm—are used as nodes, and at least one of the following characteristics—the same subject, the same shot, consecutive frames, and the same audio window—are used as edges between the nodes to construct an attribute combination graph. The attribute combination graph is matched with the protected expression graph in the preset copyright risk feature library by subgraph matching or graph vector representation similarity comparison. When the similarity between multiple nodes in the attribute combination graph and the protected expression graph in terms of combination relationship meets the preset conditions, a combination risk factor is generated. The context enhancement coefficient of the combined risk factor is determined based on at least one of the following: number of hit nodes, node rarity, relationship matching degree, and source aggregation degree. Copyright risk factors identified from different modalities are mapped to the same protected content identifier. Cross-modal consistency is determined based on the ratio of the number of modalities that hit the same protected content identifier to the total number of modalities involved in the generation conditions. The context enhancement coefficient of the corresponding copyright risk factor is adjusted based on the cross-modal consistency.
5. The copyright risk control method in the generative content reasoning process according to claim 1, characterized in that, The generation of copyright risk control instructions based on the comprehensive risk score and the copyright risk vector includes: The copyright risk vector is mapped to a copyright risk control instruction comprising at least two of the following: a mixed weight of the vector representation of the original prompt word and the vector representation of the safety prompt word; a generative model guidance strength; a negative constraint weight; a sampling trajectory offset strength; a safety prompt word; a set of negative constraints; a similarity verification threshold; and the number of local resampling times. Specifically, if the overall risk score is less than a first preset threshold, the original prompt word and reference material are retained; if the overall risk score is greater than or equal to the first preset threshold but less than a second preset threshold, the safety prompt word is generated and the negative constraint weight is increased; if the overall risk score is greater than or equal to the second preset threshold, the influence of the vector representation corresponding to the high-risk segment in the original prompt word is reduced, the influence of the vector representation of the safety prompt word and the negative constraint weight are increased, and the intermediate state of inference is reviewed. Accordingly, generating the security prompt word includes: The original prompt words are segmented into multiple semantic segments, and each semantic segment is aligned with the copyright risk factor and marked as a segment that can be retained, a segment that needs to be weakened, a segment that needs to be replaced, or a segment that needs to be deleted. The fragments to be weakened are abstracted to obtain an abstract description; Generate alternative expressions for the segments that need to be replaced, and combine the retainable segments, the alternative expressions, and the abstract descriptions into the security prompt words; Determine the semantic similarity between the security prompt word and the original prompt word after removing the weakened, replaced, or deleted segments; If the semantic similarity is less than a third preset threshold, the rewriting intensity is reduced; wherein the rewriting intensity is positively correlated with the difference between the security prompt word and the original prompt word.
6. The copyright risk control method in the generative content reasoning process according to claim 5, characterized in that, The step of controlling the generation process according to the copyright risk control instructions to obtain corresponding candidate generation results includes: In the generative content reasoning process, intermediate inference states are extracted based on a preset step size or event triggering method; the intermediate inference states include at least one of intermediate latent variables, attention maps, low-resolution preview features of the current predicted image, video keyframe features, and audio segment features; The intermediate reasoning state is compared with the preset copyright risk feature database to obtain the current intermediate risk score; Based on the relationship between the current intermediate risk score and the previous intermediate risk score corresponding to the previous review step, the sampling trajectory offset intensity corresponding to the sampling trajectory adjustment, the constraint intensity corresponding to the negative constraint enhancement, and the attention weight corresponding to the high-risk segment in the original prompt word are updated to obtain the updated copyright risk control instruction; the sampling trajectory offset intensity and the constraint intensity are parameters based on the copyright risk control instruction. According to the updated copyright risk control instructions, the process jumps back to the step of extracting intermediate states of inference based on preset step size or event triggering method until the inference is completed, so as to obtain the corresponding candidate generation results.
7. The copyright risk control method in the generative content reasoning process according to any one of claims 1 to 6, characterized in that, The process of comparing the candidate generated results based on a preset protected content library and a preset copyright risk feature library, determining a target risk score based on the comparison results, and determining a target generated result that meets a preset copyright risk threshold based on the target risk score includes: Based on the preset protected content library and the preset copyright risk feature library, the candidate generated results are compared with at least two of the following similarities: visual similarity, text semantic similarity, audio similarity, structural similarity, and entity recognition similarity, to obtain the corresponding comparison results. The similarity scores in the comparison results are weighted and summed based on the corresponding preset weights to obtain the target risk score; If the target risk score is greater than or equal to the preset copyright risk threshold, then the risk location result in the candidate generation result is determined according to the comparison result; the risk location result includes any one or a combination of several of the following: local image regions in an image, shot segments in a video, and audio windows in an audio file. Based on the risk location results, local redrawing is performed on the corresponding local image regions, segment resampling is performed on the corresponding video shot segments, or section replacement is performed on the corresponding audio windows to obtain the target generation result that meets the preset copyright risk threshold.
8. A copyright risk control device in a generative content reasoning process, characterized in that, include: The copyright risk factor identification module is used to obtain the generation conditions, preprocess the generation conditions to generate a target feature vector, and identify each copyright risk factor in the target feature vector based on a preset protected content library and a preset copyright risk feature library. The copyright risk control instruction generation module is used to encode each of the copyright risk factors to obtain a copyright risk vector, determine the comprehensive risk score corresponding to the copyright risk vector, and generate a copyright risk control instruction based on the comprehensive risk score and the copyright risk vector. The candidate generation result determination module is used to control the generation process according to the copyright risk control instructions during the generative content reasoning process in order to obtain the corresponding candidate generation results; the control operations include any one or a combination of several of the following: adjustment of the vector representation of prompt words, adjustment of attention weights, adjustment of sampling trajectories, and enhancement of negative constraints; The target generation result determination module is used to compare the similarity of the candidate generation results based on a preset protected content library and a preset copyright risk feature library, determine the target risk score according to the corresponding comparison results, determine the target generation result that meets the preset copyright risk threshold based on the target risk score, and output the corresponding copyright risk control record. The copyright risk control instruction generation module is used to map each of the copyright risk factors to a unified risk space to obtain the copyright risk vector; the copyright risk vector includes at least two risk components from the following: protected entity risk component, style risk component, composition risk component, shot risk component, text semantic risk component, audio rhythm risk component, melody risk component, reference material similarity risk component, and generation parameter risk component. The risk component is determined by multiplying the weight of the target risk dimension corresponding to each risk component, the second similarity between the target feature vector and the preset protected content library on the target risk dimension, the identification confidence, the source reliability, and the context enhancement coefficient of the copyright risk factor; wherein, the source reliability is determined according to the risk source modality of the copyright risk factor, and when there are multiple copyright risk factors in the same risk dimension, the risk component corresponding to the risk dimension is determined according to the maximum value, weighted average, or Top-K weighted aggregation method; Each risk component is weighted according to the first-order weight of the corresponding target risk dimension to obtain a first-order weighted result; the product between any two different risk components is weighted according to the corresponding interaction weight to obtain an interaction weighted result; the sum of the first-order weighted result and the interaction weighted result is normalized to obtain the comprehensive risk score.
9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program to implement the steps of the copyright risk control method in the generative content reasoning process as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, A computer program is stored on a computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the copyright risk control method in the generative content reasoning process as described in any one of claims 1 to 7.
