Video flaw traceability and fixed-point regeneration method and system

Abstract

This invention relates to the fields of artificial intelligence and digital video processing technology, and discloses a method and system for tracing the source and regenerating video defects. The method includes establishing a multi-level data lineage mapping table during the execution of the video generation pipeline, associating video frames with video segments, storyboard shots, assets, and script location layers, and carrying a set of generation parameters; receiving user-specified operations and screen coordinates for defect areas, querying the mapping table, and using a computer vision model to reverse-locate the source production stage and generation parameters; presenting an editing interface to receive parameter modifications; calculating the affected range, regenerating only the affected segments locally, and comparing the differences between the modified parameters and the original parameters to generate a change increment that is passed downstream; employing a multi-mode stitching strategy for splicing, and performing temporal smoothness and role consistency checks after stitching, switching modes or issuing alarms if unsatisfactory. This invention achieves second-level visual location of defects, reduces computational power consumption, and ensures seamless integration of repaired videos.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and digital video processing technology, specifically to a method and system for tracing the source of video defects and regenerating them at specific points. Background Technology

[0002] With the rapid development of AIGC (Artificial Intelligence Generated Content) technology, video generation technology based on large language models and diffusion models has been initially applied to the production of digital content such as comics and short dramas. Currently, the mainstream AI video generation and defect handling solutions in the industry mainly include the following categories: (1) End-to-end video generation scheme. The user inputs a text description or reference image, and the large model directly outputs a complete video clip. Typical examples include the Runway Gen series, Pika Labs, and Kling. The advantage of this type of method is its fast generation speed, but it has significant "black box" characteristics, meaning that the user cannot intervene in or control the intermediate steps of the generation process. When the generated result has flaws (such as facial distortion, limb deformation, light and shadow breaks, etc.), the user can only adjust the input parameters and regenerate the entire video, resulting in a large waste of computing power and time costs.

[0003] (2) Video defect repair solutions based on deep learning. Some researchers have proposed automatic detection and repair methods for video defects based on deep learning models. For example, the "Video Defect Removal Method Based on Deep Learning" (CN113724152A) published by Danghong Technology uses motion information estimation to automatically detect and repair video defects. However, this type of method belongs to the category of "post-repair". Its repair capability is limited by the model performance and cannot solve systemic defects caused by upstream design problems (such as improper storyboard composition and incorrect character prompts). Moreover, the repaired picture may be inconsistent with the style of the original picture, resulting in "repair traces".

[0004] (3) Multi-agent pipeline solution. Some platforms have proposed a content production pipeline solution based on multi-agent agents, which decomposes the video generation process into multiple independent stages such as script, storyboard, and video. This type of method improves the controllability of the process, but its way of handling defects is still "overall regeneration" or "manual positioning and regeneration", lacking an automated traceability mechanism from the final video to the source production stage.

[0005] (4) Frame-level metadata tracing scheme. Existing technologies include methods that embed metadata into video frames to establish a connection for tracing. However, the metadata content is mainly the node identifier of the encoding / decoding link or transmission link, rather than the lineage of generation parameters in the context of AI video generation (such as prompt words, model identifiers, random seeds, etc.). Moreover, its purpose is for content authentication or copyright tracking, rather than serving the creative iteration for defect repair.

[0006] The following technical problems are common in existing technologies: 1. Low efficiency in defect location: After discovering a defect in a video, users cannot quickly locate the source of the defect and can only check it one by one based on experience, which usually takes a long time.

[0007] 2. The repair cost is high. Overall regeneration consumes a large amount of computing power, with word consumption being dozens of times that of point-to-point repair. Traditional overall regeneration methods require re-rendering the entire video segment, while defects often only exist in local shots.

[0008] 3. Lack of interactive repair tools: Existing systems all follow a one-way process of "generate-inspect-regenerate," lacking an intuitive interactive method where users can locate and repair defects simply by double-clicking the screen. This results in a poor user experience and a high professional threshold.

[0009] 4. Repair may disrupt consistency. After partial repair, there may be inconsistencies in art style, lighting, and character appearance between the repaired and unrepaired segments, resulting in obvious "repair marks" and affecting the quality of the final product. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a method and system for tracing the source of video defects and regenerating them at specific points. This method enables automated and visual tracing of the source of video defects, provides an intuitive and interactive repair method, and offers a superior user experience. After partial repair, it ensures seamless integration with the original image while reducing the computational and time costs of defect repair.

[0011] This invention is achieved through the following technical solution: A method for tracing the source of video defects and regenerating them at specific points is provided, including the following steps: S1: During the execution of the video generation pipeline, a multi-level data lineage mapping table is established and stored; the mapping table associates video frames or frame intervals with a bottom-up source chain, which includes at least: video segment layer, storyboard layer, asset layer and script position layer; the mapping table also carries a generation parameter set, which includes at least cue words, negative cue words, model identifiers and random seeds; S2: Receive the user's specified operation on the defective area in the video playback interface, and obtain the video identifier, frame number and screen coordinates corresponding to the operation; S3: Query the mapping table based on the video identifier and frame number, and combine it with the screen coordinates to call the computer vision model to identify the image object at that coordinate. Match the identification result with the asset identifier in the asset layer to reverse locate the source production process that caused the defect and the corresponding generation parameters. S4: Presents the parameter editing interface for the source production process and receives user instructions to modify the generated parameters; S5: Based on the modified generation parameters, calculate the affected production range and regenerate only the video segments within the affected range; during the regeneration process, compare the user-modified parameters with the original parameters, generate parameter change increments, and only pass the changed parts to the downstream. Unchanged parameters are marked as inherited and the generation results of the original segments are directly reused. S6: A multi-mode stitching strategy is adopted to seamlessly stitch the regenerated video clip with the unaffected parts of the original video. The multi-mode stitching strategy adaptively selects one of the preset stitching modes to perform stitching based on the difference type between the repaired clip and the context. After stitching is completed, the stitching boundary frame is extracted to calculate the temporal smoothness, and role consistency verification is performed on the repaired area. If the smoothness is lower than the first threshold or the role consistency is lower than the second threshold, the stitching mode is automatically switched and the stitching is re-executed, or the user is prompted to adjust manually.

[0012] In this invention, during the execution of the video generation pipeline, S1 synchronously establishes and stores a multi-level data lineage mapping table. This step does not depend on user operation and is executed automatically in the system background. After S1 completes the establishment of the mapping table, it enters the repair phase, and S2 is only triggered after the video generation is completed and the user discovers defects.

[0013] S2 receives the user's specified operation (double-click / click) and obtains the video identifier, frame number, and screen coordinates. S3 uses the information obtained in S2 to query the mapping table and performs reverse localization using the CV model. After localization, it returns the source production stage and parameters. S4 presents the editing interface and receives user modification instructions. S5 receives user instructions to modify the generation parameters. S6 calculates the affected range based on the modified parameters and regenerates only the affected segments. S7 produces the regenerated video segment. S8 performs multi-mode stitching between the regenerated segment and the original video and performs post-stitching quality inspection.

[0014] The steps S1 through S6 have a clear temporal dependency. S1 is executed synchronously during the video generation pipeline, and is an accompanying behavior of the production phase. S2 through S6 are triggered sequentially after the user discovers a defect, and belong to the repair phase. Among them, S2 is the starting point of the repair phase, S3 depends on the user operation information obtained by S2, S4 depends on the location result of S3, S5 depends on the user modification command of S4, and S6 depends on the regeneration output of S5. The steps are strictly executed sequentially, with the output of the previous step serving as the input of the next step.

[0015] Furthermore, in S1, the data structure of the mapping table also includes a pixel-level defect area coordinate field, used to record the precise location range of defects in the image; and a quality metadata field, used to record the automatically detected defect type and quality score.

[0016] Furthermore, in S3, reverse positioning also includes: if the exact frame is not recorded in the mapping table, then the mapping relationship of the closest recorded frame on the timestamp is found as the approximate positioning result.

[0017] Furthermore, in S5, the calculation of the affected production scope includes: Identify all downstream stages from the initial production stage to the final video output; If the source stage is the storyboard breakdown or the generation of the first and last frames, the affected scope only includes the video segment corresponding to that shot; If the source is the asset definition stage, the affected scope includes all shots and corresponding video clips that use the asset identifier; If the source stage is the script description stage, the affected scope includes all shots and video clips corresponding to that scene; Upstream processes and parallel processes that are independent of the source process remain unchanged and are not included in the regeneration scope.

[0018] Furthermore, in S6, the adaptive selection of the multi-modal stitching strategy includes: If the clip to be repaired is located at the edge of a shot, select the hard cut mode to replace it directly without adding any transition effects; If there is a color or brightness difference between the repaired segment and the context, select the fade-in / fade-out mode and apply alpha blending transition to the first and last 1 to 8 frames of the repaired segment; If there is an action connection between the repaired segment and the context, select the optical flow interpolation mode and use the optical flow method to generate intermediate transition frames; If the appearance of a character in the repair clip is inconsistent with the context, select the feature matching mode and use face fusion or style transfer algorithms to unify the appearance features.

[0019] Furthermore, in S6, post-suture quality verification also includes: A global character consistency check is performed on the stitched complete video, the feature vectors of the characters in each frame are extracted and compared with the preset standard features; If a new visual inconsistency is detected, the location of the violation is recorded and modification suggestions are generated.

[0020] Furthermore, the method also includes a batch defect processing step: Automatically detect all defects in the video; For each defect, perform S2 to S3 to trace its origin; Defects are grouped according to their origin in the production process; For multiple defects originating from the same source, analyze their commonalities and generate a unified parameter modification scheme; Batch execute S5 to S6 to complete the repair of all defects.

[0021] This invention also provides a system using a video defect tracing and fixed-point regeneration method, comprising: The video generation pipeline module executes the complete generation process from script to video, including at least storyboard breakdown, first and last frame generation, and video clip generation functions. The data lineage tracking module records and stores the multi-level mapping relationship between video frames and each production stage in real time during the video generation process, and maintains a multi-level data lineage mapping table. The defect detection and localization module receives the user's specified operation on the defect area in the video playback interface, parses the video identifier, frame number and screen coordinates corresponding to the operation, queries the mapping table, and combines the computer vision model to reverse locate the source production link and generation parameters of the defect. The source editing module dynamically generates a parameter editing interface based on the type of source production process and accepts user modifications to the generated parameters. The incremental regeneration module calculates the production range affected by parameter modifications, calls the video generation model to regenerate only the video segments within the affected range, compares the user-modified parameters with the original parameters, generates parameter change increments, and only passes the changed parts to the downstream. The video stitching module employs a multi-mode stitching strategy to seamlessly stitch the regenerated video clips with the unaffected parts of the original video.

[0022] Furthermore, a system using video defect tracing and fixed-point regeneration methods also includes: The impact range calculation unit determines the list of processes that need to be re-executed, the list of shots that need to be regenerated, and the list of video clips that need to be regenerated, based on the source process type and the range of modified parameters. The parameter difference propagation unit compares the user-modified parameters with the original parameters, generates parameter change increments, and only propagates the changed parts to the downstream; the unchanged parameters are marked as inherited. The mapping table update unit updates the mapping records corresponding to the affected segments in the mapping table after the incremental regeneration is completed. The batch defect processing module automatically detects all defects in the video, groups the defects according to their source, generates a unified parameter modification scheme for multiple defects from the same source, and performs incremental regeneration in batches.

[0023] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables video defect tracing and point-to-point regeneration.

[0024] The beneficial effects of this invention are: I. Achieving Visual and Precise Location of Defects This invention synchronously establishes and stores a multi-level data lineage mapping table during the execution of the video generation pipeline. This mapping table associates video frames or frame intervals with a bottom-up source chain (video segment layer - storyboard layer - asset layer - script position layer) and carries a complete set of generation parameters (cue words, negative cue words, model identifiers, random seeds, etc.). When a user double-clicks on a defective area in the video frame, the system queries the mapping table based on the video identifier and frame number, and uses a computer vision model to identify the image object at that coordinate, matching the identification result with the asset identifier in the asset layer, thereby reversibly locating the source production stage that caused the defect and the corresponding generation parameters. Through the automated data lineage mapping table and pixel-level precise positioning, users can obtain the source location result within seconds after double-clicking on the defective frame. The positioning accuracy is high, transforming "manual inspection based on experience" into "system-automated visual positioning," significantly reducing the professional threshold for video content production, and enabling non-professional users to quickly locate and repair defects.

[0025] II. Significantly reduce the computational power consumption for repairs and improve production efficiency. This invention employs a parameter incremental propagation and local regeneration mechanism. After a user modifies the source parameters, the system calculates the affected production range and regenerates only the video segments within that range. During the regeneration process, the modified parameters are compared with the original parameters to generate parameter change increments. Only the changed parts are propagated downstream; unchanged parameters are marked as "inherited" and the generation results of the original segments are directly reused, avoiding the waste of computational power in overall regeneration. The repair loop time is reduced from several hours in traditional solutions to several minutes, significantly improving efficiency, and is particularly suitable for frequent iterative modifications in large-scale production scenarios such as long videos and multi-episode animated series.

[0026] III. Provides an intuitive and interactive repair experience This invention allows users to specify defective areas by double-clicking in the video playback interface. The system automatically locates the source and presents the parameter editing interface of the original production stage. The editing interface is dynamically generated according to the source type (a prompt word editor is presented in the storyboard first and last frame generation stage, a character attribute panel is presented in the asset definition stage, and a text editor is presented in the script description stage). Users do not need to understand the underlying technical details to complete the repair. Moreover, users do not need a professional technical background; the repair can be completed in just three steps: "double-click the defect - modify parameters - confirm repair." This significantly lowers the professional threshold for video content production, changes the existing system's one-way process of "generate - check - regenerate," and provides a complete interactive closed loop of "discovery - location - modification - output," significantly improving the user experience.

[0027] IV. Ensure seamless transitions and visual consistency in the restored video. This invention employs a multi-mode stitching strategy. Based on the type of difference between the repaired segment and its context (shot boundaries, color / brightness differences, motion transitions, character appearance inconsistencies), it adaptively selects one of the following modes for stitching: hard cut mode, fade-in / fade-out mode, optical flow interpolation mode, or feature matching mode. After stitching is completed, the stitching boundary frames are extracted to calculate temporal smoothness, and character consistency is checked at the repaired area. If the smoothness is lower than a first threshold or the character consistency is lower than a second threshold, the stitching mode is automatically switched and stitching is re-executed, or the user is prompted to manually adjust.

[0028] By employing an adaptive multi-mode stitching strategy, the repaired segments maintain consistency with the original video in terms of visual style, character appearance, and motion coherence, eliminating repair traces. The post-stitching quality inspection loop further ensures the repair quality, resulting in high naturalness of the transitions and reduced visibility of stitching boundaries.

[0029] V. Supports batch processing and intelligent suggestions, adapting to large-scale production. This invention supports a batch defect processing mode. The system automatically detects all defects in the video, traces the source of each defect, groups defects according to the original production stage, analyzes the commonalities of multiple defects from the same source and generates a unified parameter modification scheme, and performs incremental regeneration in batches to shorten the total repair time. The intelligent suggestion function further reduces the user's decision-making cost and improves the repair efficiency, making it particularly suitable for professional production scenarios with high requirements for final product quality.

[0030] This invention solves the core pain points of existing video generation methods, such as low defect location efficiency, high repair cost, lack of interactive tools, and poor consistency after repair, by organically combining multi-level data lineage mapping tables, interactive visual tracing, pixel-level CV precise positioning, parameter incremental transfer and local regeneration, multi-mode adaptive stitching and quality inspection closed loop. It has achieved significant results in improving positioning efficiency, reducing computing power consumption, improving user experience, and ensuring repair quality. Attached Figure Description

[0031] Figure 1 This is a flowchart of the video defect tracing and fixed-point regeneration method of the present invention.

[0032] Figure 2 This is an overall architecture diagram of a video defect tracing and fixed-point regeneration method system. Detailed Implementation

[0033] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0034] Example 1: Complete Defect Tracing and Targeted Repair Process This embodiment uses a character's facial distortion flaw appearing in the second scene of the first episode of a comic book series as an example to demonstrate the complete implementation process, such as... Figure 1 As shown.

[0035] Initial conditions: The user has completed the entire process of generating the video from the script and obtained a 30-second animated short; during the preview, the user discovered that the protagonist "Zhang San's" face was severely distorted at the 3-second mark.

[0036] (a) Establishing a mapping table (production stage) During video generation, the system automatically records mapping relationships. When a video clip is generated, the system records the mapping between the clip and the storyboard shot ID; when the first and last frames are generated, the system records the mapping between the first and last frames and the character asset ID and scene asset ID; when resource assets are registered, the system records the mapping between the asset ID and the script position.

[0037] The mapping table's data structure contains the following fields: The mapping table includes a mapping identifier (UUID), video identifier, frame sequence range, source chain (containing multi-level stage identifiers and types), generation parameter set (including prompt words, negative prompt words, model identifier, random seed, etc.), and optional pixel-level defect area coordinate fields and quality metadata fields. As an optional implementation, the mapping table can employ a hybrid storage strategy: frequently queried mapping relationships (such as the mapping from video clips to storyboard shots) are stored in a memory cache with a query latency of less than 10 milliseconds; complete mapping chains are stored in a relational database, supporting historical tracing; and fine-grained coordinate-level mappings are stored in a graph database, supporting spatial queries.

[0038] (ii) User double-click defects (interaction phase) The front-end player captures user double-click events. When a user double-clicks a defective area in the video frame, the front-end obtains the normalized coordinates of the mouse click location and the frame number corresponding to the current playback time (assuming a frame rate of 24fps), and then sends the video identifier, frame number, and coordinate information to the back-end service.

[0039] (III) Backend traceability (query stage) Upon receiving the request, the backend queries the mapping table based on the video identifier and frame sequence number to obtain the corresponding video segment identifier. If the exact frame is not recorded in the mapping table, it searches for the mapping relationship of the closest recorded frame on the timestamp. Then, it traces upwards layer by layer along the source chain to obtain the storyboard shot identifier, asset identifier, and script location in sequence. Simultaneously, combined with the screen coordinates clicked by the user, it calls a computer vision model to identify the image object (such as a face, object, etc.) at those coordinates, matches the identification results with the asset identifier, and returns the source information corresponding to the asset with the highest matching degree.

[0040] After the query is completed, the backend returns the source production process identifier, source data content, and repair suggestions, and the frontend immediately pops up the source production process editing panel.

[0041] (iv) User modification of parameters (editing stage) The front-end presents an editing interface, which is dynamically generated based on the source type.

[0042] If the source is the generation stage of the first and last frames of a storyboard, a prompt editor will be displayed, allowing users to modify the prompt text, negative prompts, sampling parameters, etc. For example, users can add the description "natural facial expressions, avoid distortion" to the prompts and strengthen the negative prompts "blurred, distorted, deformed, and superfluous limbs".

[0043] If the source is the asset definition stage, a character attribute panel will be displayed, allowing users to modify appearance tags, clothing styles, etc.

[0044] If the source is the script description section, a text editor will be displayed, allowing users to modify dialogue or scene descriptions.

[0045] (v) Incremental regeneration (regeneration stage) After the user confirms the changes, the system executes the incremental regeneration process.

[0046] First, the system calculates the affected segment range based on the source stage identifier and task ID. If the source stage is a storyboard breakdown or first / last frame generation stage, the affected range only includes the video segment corresponding to that shot. If the source stage is an asset definition stage, the affected range includes all shots using that asset identifier and their corresponding video segments. If the source stage is a script description stage, the affected range includes all shots and video segments corresponding to that scene. Upstream stages and parallel stages that are not dependent on the source stage remain unchanged and are not included in the regeneration range.

[0047] Then, the system obtains the original generation parameters for each video segment within the affected area, compares the user-modified parameters with the original parameters, generates parameter change increments, and only transmits the changed parts to the video generation model. Unchanged parameters are marked as "inherited," and the generation results of the original segments are directly reused. The system calls the video generation model to regenerate only the affected segments.

[0048] Finally, the system seamlessly stitches the regenerated segments with the unaffected parts of the original video. After stitching, the relevant information in the mapping table is updated, and the repaired complete video is output.

[0049] (vi) Delivery of Results Users receive the restored video, where flaws have been eliminated and the repaired area seamlessly integrates with the original video.

[0050] Effect verification: For a 10-minute video clip, the traditional solution requires the entire clip to be regenerated after a flaw is found, which takes an average of about 2 hours (including regeneration and manual inspection).

[0051] This invention, through a reverse positioning and pinpoint repair mechanism, reduces the average closed-loop time from the user's double-click on the defect to the output of the repaired video to 4.5 minutes, improving efficiency by more than 25 times. In 50 tests, the defect positioning accuracy reached 93%, and the smoothness of the transition between the repaired video and the unaffected parts scored 4.6 out of 5.

[0052] Example 2: Multimodal suturing strategy and quality control closed loop This embodiment details the adaptive selection of multi-mode suturing strategies and the specific implementation of the post-suturing quality inspection closed loop.

[0053] (a) Adaptive selection of suture pattern The system adaptively selects the suturing mode based on the type of difference between the repaired fragment and the context: If the clip to be repaired is located at the edge of a shot, select the hard cut mode to replace it directly without adding a transition effect; If there is a color or brightness difference between the repaired segment and the context, select the fade-in / fade-out mode and apply alpha blending transition to the first and last 1 to 8 frames of the repaired segment. If the repaired segment has action continuity with the context, select the optical flow interpolation mode and use the optical flow method to generate intermediate transition frames to ensure motion continuity; If the appearance of a character in the repair clip is inconsistent with the context, select the feature matching mode and use face fusion or style transfer algorithms to unify the appearance features.

[0054] (ii) Closed-loop quality inspection after suturing After stitching is completed, the system performs quality verification. First, several frames before and after the stitching boundary are extracted, and the temporal smoothness at the boundary is calculated. The first threshold for temporal smoothness can be set according to the video frame rate and motion complexity, with a typical value range of 0.8 to 0.95 (normalized value). In this embodiment, it is set to 0.85. When the temporal smoothness at the stitching boundary is lower than 0.85, it is determined to be a poor stitching, triggering a switch to a different stitching mode or an alarm.

[0055] Then, a character consistency check is performed on the stitched complete video. Feature vectors of the characters in each frame are extracted and compared with preset standard features. The second threshold for character consistency uses cosine similarity, typically ranging from 0.6 to 0.8; in this embodiment, it is set to 0.7. When the cosine similarity between the feature vector of the repaired character and the standard feature vector is less than 0.7, it is determined that the character appearance is inconsistent, triggering a switch to the stitching mode or an alarm. The above threshold can be calibrated using an offline test dataset or adjusted by the user according to quality requirements.

[0056] If the smoothness is below the first threshold or the role consistency is below the second threshold, the stitching mode will be automatically switched and the stitching will be re-executed, or the user will be prompted to adjust it manually.

[0057] Effect verification: In 50 tests, after adopting the adaptive multi-mode stitching strategy, the naturalness score of the connection between the repaired segment and the original video improved from 3.2 points (fixed mode stitching) to 4.6 points (adaptive multi-mode stitching), and the visibility of the stitching boundary decreased from 28% to 4%.

[0058] The specific values ​​of the first threshold (temporal smoothness) and the second threshold (role consistency) in this invention can be adjusted according to the application scenario and user needs. Under the default configuration, the first threshold is 0.85 and the second threshold is 0.7. Users can also customize the thresholds in the system settings to suit different quality requirements. The size of the threshold directly affects the repair quality: a higher threshold requires stricter suture quality but may increase the frequency of switching suture modes; a lower threshold results in faster repair speed but may tolerate minor joint imperfections.

[0059] Example 3: Batch Defect Handling and Intelligent Suggestions This embodiment provides a specific implementation method for batch defect processing.

[0060] When multiple defects exist in a video, the system supports batch processing. First, it automatically detects all defects in the video, then traces the source of each defect, grouping them according to their origin in the production process. For multiple defects from the same source, it analyzes their commonalities and generates a unified parameter modification plan. If the confidence level of the commonality is higher than a preset threshold (e.g., 0.8), batch repair is automatically performed; if it is lower than the threshold, the system prompts the user for manual review. During batch repair, the system performs parameter modification and incremental regeneration on all defects belonging to the same source at once, avoiding redundant work.

[0061] Effect verification: In a test scenario containing 10 defects, the batch processing mode reduced the total repair time from 45 minutes per repair (10 repairs x 4.5 minutes) to 12 minutes (one batch process), improving efficiency by more than 70%.

[0062] Example 4: Lightweight Implementation for Mobile Devices For resource-constrained mobile scenarios, this invention provides a lightweight implementation solution. It employs a simplified mapping table caching strategy, caching only a limited number of recent mapping relationships (e.g., the most recent 1000 mappings). Using an edge computing model, it prioritizes querying the local cache; if the cache misses, it requests simplified mapping data from the cloud (transmitting only necessary fields, such as source link identifiers and prompt word fragments). This solution reduces data transmission volume and computational resource consumption while ensuring core functionality.

[0063] Effect verification: In a mobile testing environment, the lightweight solution has an average tracing response time of 1.2 seconds, which is slightly longer than the full solution's 0.8 seconds, but reduces data transmission by about 60%, making it suitable for 4G / 5G network environments.

[0064] Example 5:

[0065] like Figure 2 As shown, a system using a video defect tracing and fixed-point regeneration method includes: The video generation pipeline module executes the complete generation process from script to video, including at least storyboard breakdown, first and last frame generation, and video clip generation functions. The data lineage tracking module records and stores the multi-level mapping relationship between video frames and each production stage in real time during the video generation process, and maintains a multi-level data lineage mapping table. The defect detection and localization module receives the user's specified operation on the defect area in the video playback interface, parses the video identifier, frame number and screen coordinates corresponding to the operation, queries the mapping table, and combines the computer vision model to reverse locate the source production link and generation parameters of the defect. The source editing module dynamically generates a parameter editing interface based on the type of source production process and accepts user modifications to the generated parameters. The incremental regeneration module calculates the production range affected by parameter modifications, calls the video generation model to regenerate only the video segments within the affected range, compares the user-modified parameters with the original parameters, generates parameter change increments, and only passes the changed parts to the downstream. The video stitching module employs a multi-mode stitching strategy to seamlessly stitch the regenerated video clips with the unaffected parts of the original video.

[0066] The system also includes: The impact range calculation unit determines the list of processes that need to be re-executed, the list of shots that need to be regenerated, and the list of video clips that need to be regenerated, based on the source process type and the range of modified parameters. The parameter difference propagation unit compares the user-modified parameters with the original parameters, generates parameter change increments, and only propagates the changed parts to the downstream; the unchanged parameters are marked as inherited. The mapping table update unit updates the mapping records corresponding to the affected segments in the mapping table after the incremental regeneration is completed. The batch defect processing module automatically detects all defects in the video, groups the defects according to their source, generates a unified parameter modification scheme for multiple defects from the same source, and performs incremental regeneration in batches.

[0067] Example 6:

[0068] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for tracing video defects and regenerating fixed-point images.

[0069] Of course, the above description is not limited to the examples given above. Technical features not described in this invention can be implemented using existing technologies, and will not be elaborated upon here. The above embodiments are only used to illustrate the technical solutions of this invention and are not intended to limit the invention. Any changes, modifications, additions, or substitutions made by those skilled in the art within the scope of this invention should be considered within the protection scope of this invention.

Claims

1. A method for tracing the source of video defects and regenerating them at specific locations, characterized in that: Includes the following steps: S1: During the execution of the video generation pipeline, a multi-level data lineage mapping table is established and stored; The mapping table associates video frames or frame ranges with a bottom-up source chain, which includes at least the following layers: video clip layer, storyboard layer, asset layer, and script location layer. The mapping table also carries a set of generation parameters, which includes at least cue words, negative cue words, model identifiers, and random seeds. S2: Receive the user's specified operation on the defective area in the video playback interface, and obtain the video identifier, frame number and screen coordinates corresponding to the operation; S3: Query the mapping table based on the video identifier and frame number, and combine it with the screen coordinates to call the computer vision model to identify the image object at that coordinate. Match the identification result with the asset identifier in the asset layer to reverse locate the source production process that caused the defect and the corresponding generation parameters. S4: Presents the parameter editing interface for the source production process and receives user instructions to modify the generated parameters; S5: Based on the modified generation parameters, calculate the affected production range and regenerate only the video segments within the affected range; during the regeneration process, compare the user-modified parameters with the original parameters, generate parameter change increments, and only pass the changed parts to the downstream. Unchanged parameters are marked as inherited and the generation results of the original segments are directly reused. S6: A multi-mode stitching strategy is adopted to seamlessly stitch the regenerated video clip with the unaffected parts of the original video. The multi-mode stitching strategy adaptively selects one of the preset stitching modes to perform stitching based on the difference type between the repaired clip and the context. After stitching is completed, the stitching boundary frame is extracted to calculate the temporal smoothness, and role consistency verification is performed on the repaired area. If the smoothness is lower than the first threshold or the role consistency is lower than the second threshold, the stitching mode is automatically switched and the stitching is re-executed, or the user is prompted to adjust manually.

2. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: In S1, the data structure of the mapping table also includes a pixel-level defect area coordinate field, which is used to record the precise location range of defects in the image; and a quality metadata field, which is used to record the automatically detected defect type and quality score.

3. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: In S3, reverse positioning also includes: if the exact frame is not recorded in the mapping table, then the mapping relationship of the closest recorded frame on the timestamp is found as the approximate positioning result.

4. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: In S5, the calculation of the affected production scope includes: Identify all downstream stages from the initial production stage to the final video output; If the source stage is the storyboard breakdown or the generation of the first and last frames, the affected scope only includes the video segment corresponding to that shot; If the source is the asset definition stage, the affected scope includes all shots and corresponding video clips that use the asset identifier; If the source stage is the script description stage, the affected scope includes all shots and video clips corresponding to that scene; Upstream processes and parallel processes that are independent of the source process remain unchanged and are not included in the regeneration scope.

5. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: In S6, the adaptive selection of multi-modal stitching strategies includes: If the clip to be repaired is located at the edge of a shot, select the hard cut mode to replace it directly without adding any transition effects; If there is a color or brightness difference between the repaired segment and the context, select the fade-in / fade-out mode and apply alpha blending transition to the first and last 1 to 8 frames of the repaired segment; If there is an action connection between the repaired segment and the context, select the optical flow interpolation mode and use the optical flow method to generate intermediate transition frames; If the appearance of a character in the repair clip is inconsistent with the context, select the feature matching mode and use face fusion or style transfer algorithms to unify the appearance features.

6. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: In S6, post-suture quality verification also includes: A global character consistency check is performed on the stitched complete video, the feature vectors of the characters in each frame are extracted and compared with the preset standard features; If a new visual inconsistency is detected, the location of the violation is recorded and modification suggestions are generated.

7. The video defect tracing and fixed-point regeneration method according to claim 1, characterized in that: It also includes batch defect handling steps: Automatically detect all defects in the video; For each defect, perform S2 to S3 to trace its origin; Defects are grouped according to their origin in the production process; For multiple defects originating from the same source, analyze their commonalities and generate a unified parameter modification scheme; Batch execute S5 to S6 to complete the repair of all defects.

8. A system using the video defect tracing and pinpoint regeneration method according to any one of claims 1 to 7, characterized in that, include: The video generation pipeline module executes the complete generation process from script to video, including at least storyboard breakdown, first and last frame generation, and video clip generation functions. The data lineage tracking module records and stores the multi-level mapping relationship between video frames and each production stage in real time during the video generation process, and maintains a multi-level data lineage mapping table. The defect detection and localization module receives the user's specified operation on the defect area in the video playback interface, parses the video identifier, frame number and screen coordinates corresponding to the operation, queries the mapping table, and combines the computer vision model to reverse locate the source production link and generation parameters of the defect. The source editing module dynamically generates a parameter editing interface based on the type of source production process and accepts user modifications to the generated parameters. The incremental regeneration module calculates the production range affected by parameter modifications, calls the video generation model to regenerate only the video segments within the affected range, compares the user-modified parameters with the original parameters, generates parameter change increments, and only passes the changed parts to the downstream. The video stitching module employs a multi-mode stitching strategy to seamlessly stitch the regenerated video clips with the unaffected parts of the original video.

9. The system according to claim 8, characterized in that, Also includes: The impact range calculation unit determines the list of processes that need to be re-executed, the list of shots that need to be regenerated, and the list of video clips that need to be regenerated, based on the source process type and the range of modified parameters. The parameter difference propagation unit compares the user-modified parameters with the original parameters, generates parameter change increments, and only propagates the changed parts to the downstream; the unchanged parameters are marked as inherited. The mapping table update unit updates the mapping records corresponding to the affected segments in the mapping table after the incremental regeneration is completed. The batch defect processing module automatically detects all defects in the video, groups the defects according to their source, generates a unified parameter modification scheme for multiple defects from the same source, and performs incremental regeneration in batches.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When a computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images