A method and system for generating an intelligent clothing pattern based on designer acceptance

Abstract

The application discloses a kind of intelligent clothing pattern generation method and system based on designer acceptance, comprising: by real-time acquisition designer multi-dimensional behavior feedback to AI-generated clothing design, establish lightweight heuristic acceptance score model, dynamic quantification designer acceptance degree to system generation result;Combined with the information of different stages of design process, the system maps acceptance and stage label to the key control parameters of the generation model, achieving diversified output adjustment for creative progress;Based on the improved Stable Diffusion generation engine, the system uses the architecture combining frozen base network and trainable copy, injects dynamic control signal through zero convolution layer, flexibly adjusts the style complexity, detail richness and fidelity of generated image.The application effectively improves the interactive efficiency and personalized adaptation capability of AI-assisted design.

Description

Technical Field

[0001] This invention belongs to the field of intelligent clothing pattern generation technology, and particularly relates to an intelligent clothing pattern generation method and system based on designer acceptance. Background Technology

[0002] With the development of AI-generated models, especially the widespread application of diffusion models, generative adversarial networks, and large-scale text-image pre-trained models (such as DALL·E and Stable Diffusion) in the field of image generation, initial explorations of AI-generated image-assisted design have begun in the field of fashion design. These systems primarily rely on text-driven image generation networks to provide designers with garment sketches, renderings, and even finished product displays. However, in actual design processes, designers' acceptance of AI-generated content often exhibits significant individual differences and periodic fluctuations, influenced by the current creative task, their personal style, aesthetic preferences, and subjective judgments about the system's output quality. Existing AIGC garment image generation methods cannot perceive and adapt to changes in designer acceptance, leading to results that deviate from design preferences and severely limiting creative efficiency. Therefore, a mechanism is urgently needed to continuously perceive designer behavioral feedback and adjust generation strategies accordingly to achieve a higher level of human-machine collaboration. In this context, introducing designer acceptance as a core indicator for regulating the generation process can effectively capture designers' acceptance and willingness to use the system-generated content at different design stages.

[0003] Most mainstream AI clothing generation systems are based on large open-source models (such as Stable Diffusion, MidJourney, and DALL·E). Their underlying structure includes a text encoder (such as CLIP), an image encoder (such as VAE), and an image generator based on U-Net or Transformer architectures. Users input free-text prompts to drive the image synthesis process. The model first maps the prompts to semantic embeddings, then gradually generates latent space images through multiple rounds of diffusion sampling, and finally decodes them into high-resolution rendered images. Building on this, some systems attempt to improve the stability and style consistency of image generation through prompt optimization mechanisms. These methods do not change the generation model itself, but rather process the prompts at the input layer. For example, some studies introduce prompt enhancement techniques, automatically adding auxiliary vocabulary such as style keywords, tailoring terms, and material descriptions to the original text, making the prompts more professional and artistically guiding. Other systems introduce style templates or keyword rewriting strategies, using designer-preset style structures or design tag libraries to structurally reorganize the original prompts, thereby improving the consistency and diversity of the generated output. These methods primarily rely on natural language processing capabilities, often combining LLM or semantic graphs from the design domain for semantic reconstruction, suitable for generating initial style drafts or rapid design brainstorming stages. Another approach improves the personalization of the generated content by introducing user feedback mechanisms. Specifically, such systems not only receive prompt input but also collect explicit or implicit behavioral data from users during use, such as whether an image is saved, deleted, or used as a basis for subsequent design evolution, as well as the time spent on the image and click behavior. This interactive data is used as feedback input to dynamically adjust the direction of subsequent image generation. Some systems use heuristic scoring methods to build simple preference models, while other research attempts to construct more complex user interest prediction mechanisms through collaborative filtering, reinforcement learning, or graph neural networks. Furthermore, to further enhance the fine-grained control of image generation, some systems employ control generation mechanisms based on structure-guided graphs. Control flow modeling methods, represented by structures such as ControlNet and T2I-Adapter, typically introduce additional inputs such as edge maps, depth maps, segmentation maps, or sketches, injecting structural information into the intermediate layers of the generation backbone network through control branch networks. These methods significantly improve the structural consistency and detail preservation of clothing images, and are highly effective in tasks such as local image modification, style transfer, and clothing reconstruction.

[0004] The shortcomings of existing technologies are as follows: For image generation methods driven by text prompts, these methods rely on static natural language input for image generation, lacking the perception and modeling of designers' interactive behaviors during actual use. They cannot dynamically adjust based on designers' acceptance of the generated results or their modification behavior, resulting in generated content that fails to match the designer's true intentions in terms of style, detail, and complexity, thus reducing the effectiveness of assisted design. For methods that improve image generation through prompt optimization mechanisms, the optimization process is usually independent of the designer's real-time feedback and creation stage, lacking adaptive adjustment capabilities based on interactive behavior. This makes it difficult to achieve coordinated control between generation strategies and design progress, leading to a disconnect between assisted content and current design needs. For methods that improve the personalization of generation by introducing user feedback mechanisms, existing methods mostly remain at the level of heuristic analysis based on explicit feedback, lacking systematic modeling and dynamic quantification of acceptance, and failing to effectively embed this trust information into the control flow of the generation model, making it difficult to establish a continuously effective trust enhancement mechanism. For control generation mechanisms based on structure-guided diagrams, the control signals are usually static graphical inputs, without dynamic adjustment based on the designer's interactive behavior or creative stage. They lack semantic adaptability and behavioral linkage mechanisms, which limits their applicability and flexibility in multi-stage design processes. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes an intelligent clothing pattern generation method and system based on designer acceptance, which effectively improves the interactive efficiency and personalized adaptation capabilities of AI-assisted design, providing the clothing design industry with an intelligent, dynamic, and reliable assisted creative solution.

[0006] To achieve the above objectives, this invention provides a method for generating intelligent clothing patterns based on designer acceptance, comprising:

[0007] Real-time collection of data on designers' interactive behavior with AI-generated clothing images;

[0008] Calculate the dynamic receptivity value based on the type and duration of the interaction behavior data;

[0009] The acceptance value and the current design stage label are mapped together to the control parameters of the image generation model;

[0010] Clothing images are generated using a hybrid architecture that includes a frozen backbone network and trainable control paths, where control signals are injected into the backbone network via zero-initialized convolutional layers.

[0011] On the other hand, to achieve the above objectives, the present invention also provides an intelligent clothing pattern generation system based on designer acceptance, comprising:

[0012] The acceptance calculation module is used to collect real-time interactive behavior data of designers on AI-generated clothing images, and calculate dynamic acceptance values ​​based on the type and duration of the interactive behavior data.

[0013] The control parameter mapping module is used to map the acceptance value and the current design stage label to the control parameters of the image generation model.

[0014] A clothing image generation module is used to generate clothing images according to the control parameters, wherein the control signals are injected into the backbone network through a zero-initialized convolutional layer.

[0015] Technical effects of this invention: This invention discloses an intelligent clothing pattern generation method and system based on designer acceptance. It employs a zero-convolutional layer to dynamically inject control signals, ensuring fine-grained adjustment of the generation process while reducing model stability fluctuations by approximately 10%, thus avoiding performance degradation. Driven by acceptance, the style diversity entropy of a single batch of 8 patterns is increased by 2.1 times. The probability of users performing the "evolution" operation significantly increases in a high-trust state, indicating that the system effectively stimulates creative extension. Real-time modeling of acceptance using multi-dimensional designer behavior data significantly improves the accuracy of capturing designer preferences, increasing user satisfaction by approximately 15%-20%. Based on a continuous interactive feedback-based acceptance adjustment mechanism, the number of design iterations is significantly reduced. In the process from "sketch" to "finished image," the cross-stage retention rate of core design elements (such as silhouette / material) reaches 92%. The number of control path parameters is only 0.36 billion (accounting for 10% of the backbone network), significantly reducing training energy consumption. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 This is a flowchart illustrating an intelligent clothing pattern generation method based on designer acceptance, according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the architecture of the dynamically controllable generation model for clothing images according to an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the structure of an intelligent clothing pattern generation system based on designer acceptance, according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] like Figure 1 As shown, this embodiment provides a method for generating intelligent clothing patterns based on designer acceptance, including:

[0023] Real-time collection of data on designers' interactive behavior with AI-generated clothing images;

[0024] Calculate the dynamic receptivity value based on the type and duration of the interaction behavior data;

[0025] The acceptance value and the current design stage label are mapped together to the control parameters of the image generation model;

[0026] Clothing images are generated using a hybrid architecture that includes a frozen backbone network and trainable control paths, where control signals are injected into the backbone network via zero-initialized convolutional layers.

[0027] Furthermore, the interactive behavior data includes the designer's acceptance, evolution, saving, rejection operations, and image hovering duration for the generated image;

[0028] The process of calculating the dynamic acceptance value includes: assigning a first positive weight to the adoption operation; assigning a second positive weight to the evolution operation; assigning a third positive weight to the save operation; assigning a negative weight to the rejection operation; and converting the hovering time through a nonlinear function before participating in the calculation.

[0029] Furthermore, the process of calculating the dynamic acceptability value also includes:

[0030] The initial acceptance level is set to a neutral value of 0.5.

[0031] The acceptance value is dynamically updated based on behavioral weights;

[0032] The real-time acceptance value is limited to a preset closed interval [0.1, 1.0] and transmitted to the generation module.

[0033] Furthermore, the hybrid architecture of the image generation model includes:

[0034] A frozen pre-trained clothing image generation network, comprising a text encoder, a latent space encoder, and a U-Net denoising network;

[0035] Trainable control paths mirror part of the encoder structure of the U-Net network;

[0036] A zero-initialized convolutional layer is connected between the output of the control path and the backbone network.

[0037] Furthermore, the process of mapping control parameters includes:

[0038] When acceptance increases, reduce the intensity of semantic guidance to increase creative divergence;

[0039] When acceptance decreases, the semantic guidance strength is increased to ensure output stability;

[0040] The style complexity, detail fidelity, and number of generated images are automatically adjusted based on the design stage labels.

[0041] Furthermore, the design stage labels include five categories: inspiration generation, sketching, material and color matching, detail control, and final rendering.

[0042] Furthermore, the process of adjusting the output image includes:

[0043] When the inspiration or sketching stage is in progress and the acceptance rate is higher than the first threshold, output multiple conceptual images with divergent compositions.

[0044] When the inspiration or sketching stage is underway and the acceptance rate is below the second threshold, output a small number of well-structured reference sketches.

[0045] When the level of detail is controlled or during the final rendering stage and the acceptance level is higher than the first threshold, output a high-fidelity image.

[0046] Output a local close-up image when the level of detail is controlled or during the final rendering stage and the acceptance level is below the second threshold.

[0047] Furthermore, the training process of the trainable control path includes:

[0048] Randomly sample acceptance values ​​and design phase labels in the training batch;

[0049] Control signals are added to the freeze network via residual injection.

[0050] Minimize the noise prediction error and jointly optimize the perceptual loss and style preservation loss.

[0051] Specifically, Figure 1The intelligent garment pattern generation method and system based on designer acceptance begins with the designer inputting text prompts and selecting a creation stage. The system then generates the first batch of basic designs. The designer's interactions with these designs—including adoption, evolution, or disregard—are captured in real-time by an event-driven acceptance calculation model and quantified into a dynamic "real-time acceptance T-value." This acceptance value, along with the original prompts and creation stage information, is then input into the dynamically controllable garment image generation module. Within this module, an adaptive control engine dynamically adjusts key generation parameters such as CFG Scale based on the acceptance level. Through a U-Net structure with trainable control paths, it directly regulates the style divergence and creativity level of the generated images without altering the original prompts. Ultimately, a new design that better aligns with the designer's implicit expectations is generated and presented, awaiting the designer's next round of interaction, thus forming a continuously self-optimizing, intelligent design closed loop.

[0052] Figure 2 The architecture of the dynamically controllable generation model for clothing images takes cue words, creation stage, real-time receptivity, latent variables, and time steps as complete inputs. Within the system, configuration parameters such as the dynamic CFG Scale are calculated based on the receptivity T and stage S, thus outputting control conditions. The generation model employs a dual-path U-Net structure, where the trainable control path injects control signals into the frozen U-Net backbone through zero-convolutional layers. Ultimately, the entire denoising process simultaneously incorporates semantic guidance from cue words and style divergence control from the dynamic CFG on the latent variables, generating highly customized clothing images.

[0053] The acceptance calculation module in the system is a high-frequency, real-time event-driven model, whose input comes from the interaction data between designers and the AI-generated system. Specific input indicators include: adoption, which refers to users explicitly clicking "adopt" in multi-image output or selecting an image as the final sketch; this action is recorded by the system as a positive feedback event; evolution, which refers to users clicking "generate more similar images" or redesigning based on an image; this behavior reflects the user's affirmation of the original image's creative direction; saving, which means users add the image to their favorites, save it locally, or export it to other modules, indicating that the image has certain usage value; rejection, which means users explicitly clicking buttons such as "delete," "unsatisfied," or "exclude"; this behavior is considered a rejection of the generated image; and hovering behavior, which is the time (in seconds) the user keeps the cursor on an image, reflecting the user's level of interest in the image.

[0054] These inputs are collected in real time by the front-end interaction system and transformed into a structured event stream, with each behavior assigned a corresponding numerical weight. The system sets the initial acceptance level to T0 = 0.5 and uses the following weighted update mechanism to recursively calculate the acceptance level:

[0055]

[0056] in, This represents the weight value of the j-th discrete behavioral event, specifically set as follows: Adopt: +0.15, Evolve: +0.10, Save: +0.05, Reject: -0.08. dwell (t j ) represents the hovering duration t j The weighting function is defined as a non-linear logarithmic decreasing function:

[0057] W dwell (t) = 0.02·log2(1+t);

[0058] For example: If a user stays on a certain image for 8 seconds, that image will contribute W. dwell =0.02·log2(9)≈0.063. The system updates the acceptance T after each interaction, with the value range limited to [0.1,1.0] to avoid boundary oscillations. This acceptance value is transmitted to the generation system in real time and serves as the core control variable affecting generation parameters (such as CFG Scale, image output type, generation style range, and quantity).

[0059] The dynamically controllable generation module for clothing images is built upon the Stable Diffusion v1.5 model. This basic model consists of a frozen backbone network, including a text encoder, a latent space encoder, and a U-Net denoising main network. The text encoder uses CLIP ViT-L / 14 to convert prompt words into 768-dimensional semantic embeddings; the VAE module compresses the image to a 64×64×4 latent variable space; the core denoising module is a set of symmetric U-Net networks, containing a 4-layer downsampling encoder, a 1-layer bottleneck layer, and a 4-layer upsampling decoder. Each layer contains ResNet residual blocks and a Cross-Attention module to introduce semantic guidance.

[0060] To introduce control over acceptability, a trainable copy of U-Net was constructed, structurally mirroring the first 13 modules of the backbone network (i.e., the encoder and bottleneck layers). A 1×1 zero-convolutional layer (ZeroConv) was inserted between each module, with its parameters initially set to zero, ensuring consistent system behavior with the original model when untuned. During training, the control path introduces adjustable signals into the backbone network via ZeroConv using an "additive injection" method: the output of each module is transformed by ZeroConv and added to the input of the same layer in the backbone network, forming a fusion channel.

[0061] Assuming the prompt words remain unchanged, the system calculates the Classifier-FreeGuidance Scale (CFG) parameters during the sampling phase based on the real-time acceptability T. The mapping logic is as follows:

[0062]

[0063] The exponent of 0.75 ensures that the high-trust region is more sensitive to CFG control; that is, the higher the trust, the smaller the CFG tends to be, and the more divergent the image generation; conversely, the system tends to be conservative. This function was derived from experimental fitting, and its specific form can be adjusted according to the system optimization requirements to balance style divergence and generation stability.

[0064] In addition, the system sets five typical creative stage labels (Inspiration Generation, Sketching, Material and Color Matching, Detail Control, and Final Rendering), each stage corresponding to a different image output format, such as color sketch, line drawing sketch, material texture display, detailed partial image, or high-fidelity simulation image. The stage label and acceptance level jointly determine: the stylistic complexity of the output image (realistic / conceptual), image fidelity (blurred / high-definition), number of outputs (1 to 10), and whether style diffusion (same style / divergent) is enabled. For example, when the stage is "Sketch" and the acceptance level is low (T≈0.2), the system sets CFG to 10.2, and the output result is 2 linear sketches; while when the acceptance level increases to T≈0.9 and the user enters the "Inspiration Generation" stage, the CFG drops to 3.9, and the output may be 8 conceptual style images, with more non-linear exploration and creative leaps in composition, materials, and colors.

[0065] The final image is obtained through iterative back-engineering by a sampler (such as DDIM or PLMS), with the number of samples, image resolution, and random seed configured uniformly by the control engine. The entire generation chain is driven by a triple condition of "prompt words + receptivity + stage information," forming a closed loop for real-time, stage-adaptive, and personalized clothing image generation.

[0066] The model training phase primarily focuses on the trainable control path. The base image generation network employs the Stable Diffusion v1.5 model, with its backbone network (including the U-Net denoising module, CLIP text encoder, and VAE latent space mapper) maintaining frozen weights. The system introduces a trainable control path that is structurally identical to the U-Net backbone encoder. This path contains 12 encoder modules and 1 bottleneck module. Each module includes convolutional blocks, residual blocks, and a self-attention mechanism, with a total of approximately 360 million parameters. A 1×1 zero-convolutional layer (ZeroConv) with extremely low parameter count is applied after the output of each module, initialized to all zeros to ensure that the representational capabilities of the backbone network are not affected during the initial training phase.

[0067] The training process uses image-cue word pairs (x, y) as basic samples. Image x is first encoded into the latent variable z0 = VAE. Enc ( x The cue word y is encoded into a semantic embedding vector E by the CLIP text encoder. y During training, the system simulates the reverse process of diffusion sampling, incorporating random noise. Add to latent variables to construct noisy samples:

[0068]

[0069] Where t∈[1,T] represents the number of diffusion steps. These are the coefficients in the standard scheduling table. The U-Net backbone network for z... t The denoising operation is performed, and its input includes the noise latent variable z. t Time step t and semantic vector E y The control path embeds the acceptability T and design phase S into a control signal c, which is then injected in the following manner:

[0070]

[0071] in ControlUNet is the output of the i-th layer backbone network. i ZeroConv represents a trainable control path. i This represents the i-th zero-convolution injection layer. This structure can be viewed as a kind of "fine-tunable residual control," which controls the application of fine-grained adjustments to the branches layer by layer, while maintaining the stable expressive power of the main branch.

[0072] The system training objective is to minimize prediction noise. The difference between the noise and the true noise ∈ constitutes the basic loss function:

[0073]

[0074] in This is the network output after including the control path. To improve visual consistency and structural control stability, a perceptual loss term can be added in the later training stages. (Based on VGG feature space) and style preservation regularization terms The final total loss is:

[0075]

[0076] Where λ1 and λ2 are adjustment factors. Training uses the AdamW optimizer, with an initial learning rate of 1×10⁻⁶. -5 The gradient clipping threshold is 1.0, the batch size is usually 64, and the total number of training iterations depends on the amount of data and convergence.

[0077] To enable the control path to generalize, the receptivity T is sampled within the range of [0.1, 1.0] during training, and five stage labels S (inspiration, sketch, material, detail, and final image) are randomly combined. This allows the model to learn the correspondence between the (T, S) combination and the generated style. The final trained model can dynamically respond to any receptivity input during inference and adapt the style control parameters of the output image according to the current design flow stage, achieving adaptive image generation based on interactive feedback and design intent.

[0078] like Figure 3 As shown, this embodiment also provides an intelligent clothing pattern generation system based on designer acceptance, including:

[0079] The acceptance calculation module is used to collect real-time interactive behavior data of designers on AI-generated clothing images, and calculate dynamic acceptance values ​​based on the type and duration of the interactive behavior data.

[0080] The control parameter mapping module is used to map the acceptance value and the current design stage label to the control parameters of the image generation model.

[0081] A clothing image generation module is used to generate clothing images according to the control parameters, wherein the control signals are injected into the backbone network through a zero-initialized convolutional layer.

[0082] Specifically, the acceptance calculation module and the clothing image dynamic controllable generation module are linked together through the intermediate variable "real-time acceptance value" and jointly serve the task of generating dynamic and controllable clothing creative images.

[0083] The acceptance rate calculation module collects real-time interaction data from designers, including accept, evolve, save, reject actions, and image hover duration. This module uses a weighted mechanism to process the data: accept actions are assigned a positive weight, evolve actions a second-highest positive weight, save actions a basic positive weight, and reject actions a negative weight. Hover duration is converted using a non-linear function before being included in the calculation. The system initializes the acceptance rate to a neutral value, dynamically updates the value based on the action weights, and transmits the real-time acceptance rate value, confining it within a preset closed interval, to the generation module.

[0084] The controllable image dynamic generation module first maps real-time receptivity to key parameters: when receptivity increases, the semantic guidance strength is reduced to stimulate creative divergence; when receptivity decreases, the guidance strength is increased to ensure output stability. Simultaneously, it automatically adjusts the style complexity, detail fidelity, and generation quantity of the output image according to the design stage (inspiration generation, sketching, material and color matching, detail control, and final rendering), setting output control conditions. The image generation model is built upon a frozen pre-trained clothing image generation model, including a text encoder, a latent space encoder, and a U-Net denoising main network. To achieve dynamic control, the system adds a trainable control path—this path replicates part of the U-Net encoder structure in the basic generation network and inserts a 1×1 convolutional layer initialized to zero at the output of each layer. The control path receives real-time receptivity and design stage labels, and the generated control signal is injected into the main network through convolutional layers in a residual superposition manner, forming a hybrid architecture of "frozen backbone + adjustable copy".

[0085] Employing a highly modular and logically closed-loop structural system, its overall architecture revolves around a four-stage main process: "designer behavior perception—trust modeling—generation strategy control—image synthesis feedback." In terms of physical deployment, it adopts a decoupled front-end and back-end structure, with core control logic concentrated in the back-end generation service system, while the front-end is responsible for event collection and parameter transmission. The overall system structure can be viewed as an intelligent generation loop driven by behavior flow, with its various modules achieving asynchronous coupling and orderly scheduling through a set of structured state variables (including acceptance values, stage labels, and feedback flags).

[0086] In this structure, the acceptability modeling subsystem is located at the entry point of the overall data flow and serves as the preliminary control source for generation scheduling. This subsystem communicates in real-time with the user interaction layer, recording behavioral data and immediately updating the state upon event occurrence, generating a floating-point acceptability coefficient T∈[0.1,1.0], which serves as the master control signal for the downstream image generation control module. The acceptability state is continuously maintained by the system and pushed to the scheduling engine via message channels or an internal shared memory mechanism. Simultaneously, the system identifies the creation stage state s based on the user's current operational context. This state exists in the form of discrete classification labels, with five categories corresponding to the five typical stages in the clothing design process. Acceptability and stage labels form a set of joint control factors, constituting the current "generation intention state."

[0087] After the intent state is sent to the control mapping module, it is mapped into an image generation control package. Internal parameters include: CFGScale value, output image format instruction, number of images to be generated, whether style divergence is allowed, style scheduling coefficient, and seed sequence. This control package, along with the prompt, serves as input to the image generation subsystem. Within the image generation module, the numerical fields of this control package are decoded level by level and mapped to sub-components such as the sampler, U-Net control path, and rendering formatter. Within the U-Net layer, communication between the control path and the backbone network is achieved through Zero-Conv connections. The location and intensity of the injected control signal are dynamically determined by the real-time parameters in the control package, forming a generation control flow driven by receptivity.

[0088] Once generated, images are sent to the front end in batches for users to view, accept, or modify. The front-end event listening mechanism then feeds back all user actions on the images in this round to the system, enabling the next round of acceptance updates and stage recognition, thus forming a closed adaptive generation loop. The entire structure maintains a decoupled upstream and downstream architecture, centralized data state control, and clear parameter translation flow. It supports a complete closed loop of asynchronous state-driven batch image generation and behavior response in multi-user, multi-design task concurrent scenarios. This structure not only has clear hierarchical logic and communication channel divisions but also provides an engineering foundation for the scalability and modular parallelism of the system's subsequent deployment in a distributed architecture.

[0089] Alternative Solution 1: Employ a machine learning classification model based on interactive behavior features to model designer acceptance, replacing the original weighted scoring method. Specifically, the system no longer directly maps acceptance increases or decreases to single events. Instead, it constructs a structured feature vector from multiple behavioral data points of the user over a recent period (e.g., the last 5 interactions), inputting this vector into a pre-trained lightweight model for discrimination and output. Feature dimensions include, but are not limited to: image acceptance rate (number of acceptances in the past N rounds / total number of candidate images), average number of modifications, click-to-negation ratio, average dwell time per round, and interaction rhythm (number of operations per unit time). The model can be implemented using XGBoost or a shallow neural network, outputting an acceptance level (e.g., high, medium, low) or a normalized confidence score, which can be mapped to the generation control module via soft labels.

[0090] Alternative Solution 2: Instead of using the CFG Scale parameter for image divergence adjustment in the image generation control mechanism, a similar effect is achieved using a guided strategy based on latent space perturbation amplitude control. Specifically, a set of adjustable style perturbation vectors δ is introduced during the generation process of the StableDiffusion model. z This vector in the latent space is related to the latent variable z of the main image. t Superposition forms the synthetic latent variable z′ t=z t +α·δ z Where δ z The perturbation vector is sampled from the style distribution or generated through a style transformation network, where α is the perturbation strength coefficient, obtained by mapping from the acceptance level T. For example:

[0091]

[0092] When acceptance is high, α is larger, introducing stronger style shifts into the image generation process, resulting in more creative and unpredictable outputs. Conversely, when acceptance is low, α tends to 0, and the model behavior reverts to a stable generation state. This scheme does not rely on the CFG control path and can be applied to various Diffusion architectures, especially suitable for variant models with unstable CFG convergence or limited style control by CFG intensity, while also possessing higher internal model compatibility.

[0093] This invention discloses an intelligent clothing pattern generation method and system based on designer acceptance, aiming to solve the problem that existing AIGC clothing image generation methods cannot perceive and adapt to changes in designer acceptance, leading to results that deviate from design preferences and severely limit creative efficiency. The system establishes a lightweight heuristic acceptance scoring model by collecting real-time multi-dimensional behavioral feedback from designers on AI-generated clothing designs (including acceptance, modification, rejection, saving, and dwell time), dynamically quantifying the degree of designer acceptance of the system's generated results. Combining information from different stages of the design process, the system maps acceptance and stage labels to key control parameters of the generation model, enabling diversified output adjustments based on creative progress. Based on an improved Stable Diffusion generation engine, the system adopts an architecture combining a frozen base network and trainable copies, injecting dynamic control signals through zero-convolutional layers to flexibly adjust the style complexity, detail richness, and fidelity of the generated images, supporting multi-level design expression from inspirational sketches to high-fidelity rendering. This method effectively improves the interactive efficiency and personalized adaptation capabilities of AI-assisted design, providing the clothing design industry with an intelligent, dynamic, and reliable assisted creative solution.

[0094] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating intelligent clothing patterns based on designer acceptance, characterized in that, include: Real-time collection of data on designers' interactive behavior with AI-generated clothing images; Calculate the dynamic receptivity value based on the type and duration of the interaction behavior data; The acceptance value and the current design stage label are mapped together to the control parameters of the image generation model; The process of mapping control parameters includes: reducing semantic guidance strength to increase creative divergence when acceptability increases; increasing semantic guidance strength to ensure output stability when acceptability decreases; and automatically adjusting the style complexity, detail fidelity, and number of generated output images based on design stage labels. Clothing images are generated using a hybrid architecture that includes a frozen backbone network and trainable control paths, wherein control signals are injected into the backbone network via zero-initialized convolutional layers; The interactive behavior data includes the designer's acceptance, evolution, saving, rejection actions on the generated image, and the duration of image hovering. The process of calculating the dynamic acceptance value includes: assigning a first positive weight to the adoption operation; assigning a second positive weight to the evolution operation; assigning a third positive weight to the save operation; assigning a negative weight to the rejection operation; and converting the hovering time through a nonlinear function before participating in the calculation. The process of calculating the dynamic acceptance value also includes: The initial acceptance level is set to a neutral value of 0.

5. The acceptance value is dynamically updated based on behavioral weights; The real-time acceptance value is limited to a preset closed interval [0.1, 1.0] and transmitted to the generation module.

2. The intelligent clothing pattern generation method based on designer acceptance as described in claim 1, characterized in that, The hybrid architecture of the image generation model includes: A frozen pre-trained clothing image generation network, comprising a text encoder, a latent space encoder, and a U-Net denoising network; Trainable control paths mirror part of the encoder structure of the U-Net network; A zero-initialized convolutional layer is connected between the output of the control path and the backbone network.

3. The intelligent clothing pattern generation method based on designer acceptance as described in claim 1, characterized in that, The design stage labels include five categories: inspiration generation, sketching, material and color matching, detail control, and final rendering.

4. The intelligent clothing pattern generation method based on designer acceptance as described in claim 3, characterized in that, The process of adjusting the output image includes: When the inspiration or sketching stage is in progress and the acceptance rate is higher than the first threshold, output multiple conceptual images with divergent compositions. When the inspiration or sketching stage is underway and the acceptance rate is below the second threshold, output a small number of well-structured reference sketches. When the level of detail is controlled or during the final rendering stage and the acceptance level is higher than the first threshold, output a high-fidelity image. Output a local close-up image when the level of detail is controlled or during the final rendering stage and the acceptance level is below the second threshold.

5. The intelligent clothing pattern generation method based on designer acceptance as described in claim 2, characterized in that, The training process of the trainable control path includes: Randomly sample acceptance values ​​and design phase labels in the training batch; Control signals are added to the freeze network via residual injection. Minimize the noise prediction error and jointly optimize the perceptual loss and style preservation loss.

6. A system for an intelligent clothing pattern generation method based on designer acceptance according to any one of claims 1-5, characterized in that, include: The acceptance calculation module is used to collect real-time interactive behavior data of designers on AI-generated clothing images, and calculate dynamic acceptance values ​​based on the type and duration of the interactive behavior data. The control parameter mapping module is used to map the acceptance value and the current design stage label to the control parameters of the image generation model. A clothing image generation module is used to generate clothing images according to the control parameters, wherein the control signals are injected into the backbone network through a zero-initialized convolutional layer.

Images