Clothing pattern personalized customization system and method based on improved Stable Diffusion model

Abstract

The invention discloses a garment pattern personalized customization system and method based on an improved Stable Diffusion model, the system comprises an input module, a core processing module and an output module, the input module receives multi-modal user input and fuses the multi-modal user input; the core processing module generates an initial pattern by using an improved Stable Diffusion model, realizes accurate adaptation and parameterization control of the pattern and a clothing model through a TPS algorithm, CLIP semantic analysis, Perlin noise and the like, and reduces feedback delay in combination with 3D real-time rendering; the method comprises the steps of multi-modal pattern generation, dynamic adaptation and parameter control and production file output, the full-link bottleneck of'design-adaptation-production 'of garment customization is broken through, normal form transition from manual leading to AI driving is achieved, and the production efficiency is improved. And personalized and industrial production standards are considered.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence generated content (AIGC) and clothing design, and specifically relates to a clothing pattern individualization customization method and system based on an improved Stable Diffusion (SD) model, which realizes a one-stop digital solution from user input to flexible production. BACKGROUND

[0002] There are significant technical bottlenecks in the current clothing customization field: the traditional mode relies on manual design processes, resulting in long average development cycles and high single-cost, making it difficult to meet the needs of the modern consumer market for rapid response; although existing AI generation technology has improved efficiency, the adaptability of the generated patterns to clothing patterns (such as sleeve openings, collar openings, and other key parts) is insufficient, causing spatial misalignment problems between patterns and clothing structures. In addition, existing systems lack real-time user interaction mechanisms, and consumers cannot dynamically adjust parameters such as pattern density, color saturation, and element proportion, and the design feedback is delayed for a long time, which seriously restricts the personalized experience. More critically, the output file format of existing technologies cannot be directly interfaced with production systems, lacking the ability to generate vector files that meet industrial standards (such as CMYK color separation and cutting piece positioning), resulting in low conversion efficiency from design to manufacturing. The above defects collectively result in a technical dilemma of a "design-adaptation-production" full-link fault in the industry. SUMMARY

[0003] To solve the above-mentioned problems of the prior art, the present application provides a clothing pattern individualization customization system and method based on an improved Stable Diffusion model, which can break through the "design-adaptation-production" full-link bottleneck of clothing customization, realize a paradigm shift from human-led to AI-driven, and balance individualization and industrial production standards.

[0004] To achieve the above-mentioned purpose, the present application is realized by the following technical solutions:

[0005] The clothing pattern individualization customization system based on the improved Stable Diffusion model includes an input module, a core processing module, a 3D real-time rendering unit, and an output module, which are communicatively connected.

[0006] The input module is used to receive user input, including basic design specifications (such as clothing material, resolution requirements) and user creative descriptions (such as style theme, color preference), as well as user-uploaded sketches / reference images, for multi-modal fusion to generate text / sketch / image information.

[0007] The core processing module includes a Stable Diffusion generation unit that generates an initial garment pattern based on the multi-modal information of the input module; a dynamic adaptation engine that performs nonlinear geometric deformation on the generated pattern using a thin plate spline interpolation (TPS) algorithm based on garment pattern key points (such as sleeve cap curve and neckline vertex), so that the pattern accurately fits the garment surface;

[0008] The 3D real-time rendering unit combines parameterized control (density / color / size) with a 3D physical rendering engine (Blender Cycles) to perform real-time rendering on the deformed pattern.

[0009] The output module generates production files, including automatically generating vector files (SVG / AI) and process parameter packages (including embroidery needle count / cutting piece positioning) that meet industry standards (such as CMYK color separation and cutting piece positioning) to interface with digital direct jet machines / laser cutting machines.

[0010] Preferably, in the multi-modal fusion of the input module, the basic design specification and the user creative description are semantically fused to generate a joint prompt vector for the AIGC model (Stable Diffusion), with the formula being P fused =αP base +(1-α)P user , where P fused represents the fused joint prompt vector; P base represents the basic design specification vector (such as material and resolution); Puser represents the user creative description vector (such as style and color); and α∈[0.6, 0.8] is a specification weight coefficient that is dynamically adjusted through industry standard design rules.

[0011] Preferably, the core processing module further includes a CLIP alignment unit that maps the user-uploaded sketch / reference image to the same semantic space as the text prompt, thereby constraining the visual consistency of the generated pattern through cosine similarity and preventing inconsistencies between the image and the text (such as a user uploading a flower sketch but generating geometric stripes); through the cross-modal feature projection formula , where F image represents the image feature vector; F text represents the text prompt feature vector; and ||F|| represents the Euclidean norm of the feature vector, and the similarity threshold is >0.75 to trigger automatic optimization.

[0012] Preferably, in the TPS deformation of the dynamic adaptation engine, the energy minimization problem of the thin plate spline interpolation satisfies , where E represents the energy function of the thin plate spline interpolation; represents the second derivative of the function f with respect to x (controls bending stiffness); represents the mixed second derivative (controls shear deformation).

[0013] Preferably, the core processing module further comprises a Perlin noise unit to generate a natural transition density map through a fractal noise algorithm, avoid a mechanical repetition (such as an organic distribution of plant texture in a sweatshirt print), control random changes in the density of pattern elements, and superimpose a gradient field interpolation and an octave noise formula, which is:

[0014] wherein D(x, y) represents a density map (value range [0, 1], 0 represents no elements, and 1 represents the maximum density); i represents an octave noise layer index (i = 0, 1,..., k, and k is usually taken as 4); noise(2ix, 2iy) represents a Perlin noise value of the i-th layer (the frequency is 2 i , and the natural texture details are simulated); H represents a Hurst index (value range [0, 1], and the default is 0.7, which controls the roughness of the noise); 2 i(2-H) represents an amplitude attenuation factor, which supports real-time adjustment of the “complexity” parameter (density coefficient D ∈ [0.1, 2.0]) by the user.

[0015] The application discloses a clothing pattern personalized customization method based on an improved Stable Diffusion model, which comprises the following steps:

[0016] Multi-modal pattern generation: fusing basic design specifications and user creative description semantics to generate a joint prompt word vector, combining a user-uploaded sketch / reference picture, and generating an initial clothing pattern through CLIP alignment;

[0017] Dynamic adaptation and parameter control: based on clothing pattern key points, using a TPS algorithm to perform nonlinear geometric deformation on the initial pattern, and then controlling the density of pattern elements through a Perlin noise algorithm to output a final pattern;

[0018] Production file output: automatically generating vector files and process parameter packages in line with industrial standards to interface production equipment.

[0019] Preferably, in the multi-modal pattern generation step, the joint prompt word vector generation formula is

[0020] P fused = αP base +(1-α)P user , the CLIP alignment generates visual consistency of the pattern through cosine similarity constraint, and the formula is When the similarity threshold is greater than 0.75, automatic optimization is triggered.

[0021] Preferably, in the dynamic adaptation and parameter control step, the TPS deformation is based on a thin plate spline interpolation energy minimization formula:

[0022] Nonlinear geometric deformation is performed, Perlin noise is interpolated through a gradient field, and an octaves noise superposition formula is used: The density of the pattern elements is controlled, and the user can adjust the density coefficient D in the range of [0.1, 2.0].

[0023] Beneficial effects: The present application realizes the paradigm shift from manual guidance to AI driving in clothing customization, solves the whole-linkage collaboration problem of design-adaptation-manufacturing, and balances personalized needs and industrial production standards; The present application breaks through the whole-linkage bottleneck of "design-adaptation-production" in clothing customization, realizes the paradigm shift from manual guidance to AI driving, balances personalized and industrial production standards, specifically supports multi-modal input of text, sketch and image by using an improved StableDiffusion model, realizes rapid pattern generation, and replaces the traditional manual design process; Based on the thin plate spline interpolation algorithm and the CLIP semantic analysis technology, the geometric mapping relationship between the pattern and the garment pattern is established, and accurate matching is realized; The user adjustment feedback delay is reduced by integrating parameterized control and 3D physical rendering engine; CMYK color separation vector files and process parameters are automatically generated, and are directly connected to digital direct jet machines / laser cutting machines. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is the system architecture diagram of the present application.

[0025] Figure 2 is the dynamic adaptation algorithm flowchart of the present application.

[0026] Figure 3 is the user interaction interface schematic diagram of the present application.

[0027] Figure 4 is the implementation flowchart of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0029] Overall scheme / principle of the present application:

[0030] The application constructs a set of garment pattern customization system integrating AIGC generation, intelligent adaptation and real-time interaction, and breaks through the industry bottleneck through the following technical solutions: an improved StableDiffusion model is used to support multi-modal input of text, sketch and image, to realize rapid pattern generation and replace the traditional manual design process; a geometric mapping relationship between the pattern and the garment pattern (neck opening / sleeve opening, etc.) is established based on the thin plate spline interpolation (TPS) algorithm and the CLIP semantic analysis technology, to realize accurate matching; a parameterized control system (density / color / size) and a 3D physical rendering engine (BlenderCycles) are integrated to reduce the user adjustment feedback delay; CMYK color separation vector files (SVG / AI) and process parameter packages (including embroidery needle number / cutting piece positioning) are automatically generated to directly interface with digital direct jet machines / laser cutting machines. The application realizes the paradigm shift of garment customization from "manual dominant" to "AI driven", solves the whole-link collaborative problem of design-adaptation-manufacturing, and balances the personalized demand and the industrial production standard.

[0031] I. Input module: receives the basic design specifications (such as the requirement of cotton fabric and the resolution of 300 dpi) input by the user, the user's creative description (such as the desire for a retro style and a pattern dominated by blue), and the sketch and reference picture uploaded by the user. Multi-modal fusion is performed to fuse the semantic of the basic design specification and the user's creative description to generate a joint prompt word vector of the AIGC model, and the formula is P fused =αP base +(1-α)P user , where a is dynamically adjusted in [0.6, 0.8] to ensure that the generated pattern meets the production specification (such as avoiding the hollow area exceeding the cutting piece boundary) and retains the core elements of the user's creative idea. At the same time, the sketch / reference picture uploaded by the user is aligned and mapped to the same semantic space as the text prompt word through CLIP, and the cosine similarity formula is used to constrain visual consistency, and when the similarity threshold is greater than 0.75, it is automatically optimized to prevent "text and picture inconsistency".

[0032] II. The core processing module includes a StableDiffusion generation unit and a 3D real-time rendering unit

[0033] 1. StableDiffusion generation unit: according to the multi-modal information of the input module, an initial garment pattern is quickly generated to replace the traditional manual design and improve the design efficiency.

[0034] 2. Dynamic adaptation engine: based on the key points of the garment pattern (such as the bending point of the sleeve hole arc line and the position of the neck opening vertex), the TPS algorithm is used for nonlinear geometric deformation. The energy minimization formula is

[0035] Let the generated pattern fit the garment curve accurately, solve the distortion problem caused by pattern stretching / compression (such as T-shirt chest pattern, can accurately match the chest curvature). Also contains Perlin noise unit, through fractal noise algorithm, with formula Generate natural transition density distribution, control the random density change of pattern elements, avoid mechanical repetition (such as plant texture in sweatshirt printing, can have organic distribution), support user to adjust density coefficient D∈[0.1,2.0].

[0036] 3.3D real-time rendering unit: combine parameterized control system (user can adjust density, color, size, etc.) and 3D physical rendering engine (BlenderCycles), real-time render the deformed pattern, let the user can directly see the adjusted effect, reduce feedback delay.

[0037] Three, output module: automatically generate vector files (SVG / AI format, including CMYK separation, cutting positioning information) and process parameter package (such as embroidery needle number, cutting specific positioning data, etc.) that meet industrial standards, directly interface with digital direct jet machine, laser cutting machine and other production equipment, realize efficient conversion from design to production.

[0038] Embodiment

[0039] (I) Multi-modal pattern generation

[0040] Dual prompt word fusion: fuse the basic design specification (garment material, resolution requirement, etc.) and user creative description (style theme, color preference, etc.) semantics, according to formula:

[0041] P fused =αP base +(1-α)P user Generate joint prompt word vector to control StableDiffusion model, ensure that the generated pattern meets production specifications and retains user creativity.

[0042] CLIP alignment: map the user uploaded sketch / reference map to the text prompt word semantic space, use cosine similarity formula to constrain the visual consistency of the generated pattern, automatically optimize when the similarity threshold>0.75, prevent "text and picture do not match", quantify the evaluation of generation quality.

[0043] (II) Dynamic adaptation and parameter control

[0044] TPS deformation: based on garment pattern key points (sleeve collar arc line, collar vertex, etc.), use TPS algorithm for non-linear geometric deformation of generated pattern, according to the energy minimization formula to make the pattern fit the garment curve accurately, solve the distortion problem.

[0045] Perlin noise: A density distribution map is generated using a fractal noise algorithm. The density of pattern elements is controlled by gradient field interpolation and octave noise superposition formula. Users can adjust the "complexity" parameter (density coefficient D∈[0.1,2.0]) to avoid mechanical repetition. After the above steps, the final pattern is output.

[0046] The present invention will be described in conjunction with the accompanying drawings:

[0047] like Figure 1 As shown, the input module first receives user input, obtains text / sketch / image information through multimodal fusion, and then passes it to the core processing module. The StableDiffusion generation unit of the core processing module generates the initial pattern, which is then processed by the dynamic adaptation engine (including TPS deformation) and the 3D real-time rendering unit. Finally, the output module generates production files (vector files and process sheets) for integration with production equipment.

[0048] like Figure 2 As shown, the dynamic adaptation algorithm process first performs key point detection for the garment pattern; then, CLIP semantic segmentation is performed on the generated pattern. The results of both are used for TPS control point extraction, followed by nonlinear deformation of the pattern to obtain the adaptation result.

[0049] like Figure 3 As shown in the diagram, the user interface allows users to pass parameters to the real-time rendering unit through interactive items such as density slider, color adjustment, and size control. These parameters are achieved using techniques such as Perlin noise, LAB conversion, and non-uniform scaling, ultimately enabling 3D preview and facilitating real-time adjustments and viewing of the effects.

[0050] like Figure 4 The diagram shown illustrates the multimodal pattern generation, dynamic adaptation, and parameter control process, illustrating the key steps from user input to the final usable clothing pattern output.

[0051] Dual-cue word fusion in multimodal pattern generation: Semantically fuse basic design specifications (such as production constraints like clothing material and resolution requirements) with user creative descriptions (such as personalized needs like style themes and color preferences) to generate joint cue word vectors that control the AIGC (StableDiffusion) model. This ensures that the generated patterns not only meet production standards but also retain the core of the user's creative ideas, setting the tone for subsequent pattern generation.

[0052] CLIP alignment in multimodal pattern generation: The user-uploaded sketch / reference image is mapped to the same semantic space as the text prompt. The visual consistency of the generated pattern is constrained by the cosine similarity formula to avoid "image and text mismatch" (e.g., the sketch is a flower, but the generated pattern is geometric stripes). When the similarity threshold is >0.75, it will automatically optimize and quantify to ensure the generation quality, and finally obtain the initial pattern.

[0053] TPS deformation in dynamic adaptation and parameter control: Based on garment pattern key points (such as sleeve hole arc, neckline vertex, etc.), the thin plate spline interpolation (TPS) algorithm is used to make nonlinear geometric deformation to the initial pattern, and according to the energy minimization formula, the pattern is accurately fitted to the garment curve, solving the distortion problem caused by pattern stretching / compression (such as fitting the T-shirt chest pattern to the human curve).

[0054] Perlin noise in dynamic adaptation and parameter control: Through fractal noise algorithm, a natural transition density distribution map is generated by formula to control the random density change of pattern elements, avoiding mechanical repetition (such as making the plant texture distribution of sweatshirt prints more natural), and supporting user adjustment of density coefficient to realize personalized parameter control. After this processing, the final pattern output can be used for production, connecting the subsequent production system link.

[0055] Through multi-modal input to generate an initial pattern, and then deform, noise control to adapt to garment patterns and individual needs, it is a key technical path from creative to usable results for clothing patterns, helping to realize AI-driven clothing customization.

[0056] Further, the structure of "input module, core processing module, 3D real-time rendering unit, output module", related technologies and adaptation directions are explained, but the following content should not be understood as a limitation of the technical solution. In specific implementation, it can be selected and used according to the specific scene.

[0057] I. Input module

[0058] (1) Structure and function logic

[0059] Structure: It can be composed of user interactive terminal (web page / client interface), multi-modal data acquisition unit (text analysis, image / sketch upload processing module), data preprocessing and fusion sub-module.

[0060] The interactive terminal is responsible for receiving the user input basic design specification (text description), creative description (text), sketch / reference map (image file);

[0061] The acquisition unit performs semantic word segmentation and vectorization on the text, and performs format conversion (such as converting to PNG uniform format) and resolution preliminary screening on the image;

[0062] The fusion sub-module executes the double prompt word fusion algorithm to generate a joint prompt word vector to control Stable Diffusion, and at the same time completes the pre-image feature extraction of CLIP alignment.

[0063] Technology Implementation and Adaptation Direction: If deployed on a general-purpose server (such as Dell PowerEdge series), rely on server CPU (Intel Xeon series), memory (DDR4 high-frequency memory), and Python to write interactive interfaces (Flask / Django framework) to achieve web-based interaction. Use OpenCV to process images and NLTK / Transformers library for text semantic analysis. If targeting lightweight cloud services, you can use serverless architectures such as AWS Lambda. Image / text processing can be done using AWS Rekognition and Comprehend services for preprocessing, and custom fusion logic. For consumer-grade hardware (such as high-end workstations), directly run the client program (developed with PyQt) locally, and use local GPU (NVIDIA RTX 40 series) to accelerate image feature calculation.

[0064] II. Core Processing Module

[0065] (I) Structure and Functional Logic

[0066] Structure: Can include Stable Diffusion model inference unit, dynamic adaptation engine (TPS deformation submodule, Perlin noise submodule), CLIP semantic alignment and optimization submodule.

[0067] Stable Diffusion Inference Unit: Load improved model weights, receive joint prompt words and aligned image features from input module, perform diffusion model inference, and generate initial clothing patterns.

[0068] Dynamic Adaptation Engine: TPS deformation submodule calls numerical calculation library (such as SciPy), solves thin plate spline interpolation equation based on clothing pattern key point coordinates, and completes pattern geometric deformation. Perlin noise submodule uses NumPy to implement fractal noise algorithm to generate density distribution and control pattern element density.

[0069] CLIP Semantic Alignment Submodule: Real-time calculation of cosine similarity between generated patterns and text / reference images triggers optimization logic (such as re-iteration generation, local adjustment).

[0070] Technology implementation and adaptation direction: Model inference relies on GPU acceleration, NVIDIA GPU cooperates with CUDA, cuDNN framework, loads and runs Stable Diffusion model with PyTorch; Numerical calculation (TPS deformation) requires CPU single core computing power, AMD Ryzen Threadripper series or Intel Xeon Platinum series CPU can efficiently handle matrix operations; If you pursue inference speed, you can use a dedicated AI accelerator card (such as NVIDIA A100), combined with TensorRT to quantize and optimize model inference, reduce latency.

[0071] Three, 3D real-time rendering unit

[0072] (I) Structure and functional logic

[0073] Structure: Can be composed of 3D scene construction submodules (load clothing pattern 3D model, texture mapping), parameterized control interface (receive density, color, size adjustment instructions), real-time rendering engine (such as Blender Cycles integration).

[0074] 3D scene construction: Import clothing basic 3D model (such as FBX / OBJ format), map the dynamically adapted pattern as a texture to the clothing surface; Parameterized control: Receive user interaction instructions (such as density slider, color adjustment), real-time convert to rendering parameters (such as material node attributes, texture coordinate transformation); Real-time rendering: Call Blender Cycles engine, use GPU ray tracing acceleration, quickly generate 3D preview image, feedback to user interface.

[0075] Technology implementation and adaptation direction: 3D model processing relies on Blender Python API to realize model import, texture binding and parameter driving; Real-time rendering requires GPU support for ray tracing, NVIDIA RTX series graphics card can enable OptiX acceleration; If facing web-side real-time preview, you can combine Three.js framework to migrate 3D rendering to the browser, use client GPU computing power through WebGL to reduce server pressure.

[0076] Four, output module

[0077] (I) Structure and functional logic

[0078] Structure: Can include production file generation sub-module (vector file export, process parameter packaging), production docking interface (compatible with digital direct jet machine / laser cutting machine protocol). Production file generation: Convert the final pattern into CMYK color separation SVG / AI vector file, extract process parameters (embroidery needle number, cutting positioning coordinates) to generate JSON / XML format process package; Production docking: According to different production equipment protocols (such as PDF printing protocol of digital direct jet machine, G-code instruction of laser cutting machine), convert and output the file, and directly push it to the equipment queue. Technical implementation and adaptation direction: Vector file generation uses the script interface (such as Python calling Inkscape command line) of Inkscape or Illustrator to realize pattern vectorization and color mode conversion; Process parameter extraction relies on the parameter recording of the TPS deformation and Perlin noise links in the early stage, and integrates to generate structured data; Production docking needs to adapt to the device API, such as connecting Mimaki digital direct jet machine, uploading the printing task through its SDK.

[0079] In this application, the configuration of "input module, core processing module, 3D real-time rendering unit, output module" needs to be combined with software and hardware (such as GPU selection, open source library / commercial software matching) according to the scene (local workstation, cloud SaaS, large-scale factory collaboration), from the perspective of computing power, cost and compatibility, to realize the whole process from user input to production output. The core is to connect each module with algorithms (double prompt word fusion, TPS deformation, etc.), so that the AI generated clothing pattern can efficiently adapt to production needs.

[0080] Finally, it should be noted that the present application is not limited to the above embodiments, but can have many variations. All variations that can be directly derived or inferred from the disclosed content by those of ordinary skill in the art should be considered within the scope of the present application.

Claims

1. A clothing pattern personalization customization system based on an improved Stable Diffusion model, characterized in that, The input module, the core processing module, the 3D real-time rendering unit output module, and the communication connection are included. The input module is used for receiving user input, which covers basic design specifications and user creative descriptions, as well as user-uploaded sketches and / or reference images, performing multi-modal fusion, and generating multi-modal information; the design specifications include clothing material and resolution, and the user creative descriptions include style theme and color preference; the multi-modal information includes text, sketch, and image information The core processing module includes a StableDiffusion generation unit for generating an initial clothing pattern based on the multi-modal information of the input module, a dynamic adaptation engine for performing nonlinear geometric deformation on the generated pattern based on clothing pattern key points using a thin-plate spline interpolation (TPS) algorithm to make it accurately fit the clothing surface, and clothing pattern key points including sleeve hole arc lines and collar vertexes. The 3D real-time rendering unit combines parameterized control and a 3D physical rendering engine to perform real-time rendering on the deformed pattern; the parameters in the parameterized control include density, color, and size. The output module is used for generating production files, including automatically generating vector files and process parameter packages that meet industrial standards, and directly inputting data into a digital direct jet machine and / or a laser cutting machine; the vector files (SVG / AI) include CMYK separation and piece positioning; and the process parameter package includes embroidery needle count and piece positioning.

2. The clothing pattern personalization customization system based on the improved Stable Diffusion model according to claim 1, wherein, In the multi-modal fusion of the input module, the semantic fusion of the basic design specification and the user creative description generates a joint prompt word vector of the control AIGC model Stable Diffusion, and the formula is P fused =αP base +(1-α)P user , wherein P fused represents the fused joint prompt word vector; P base represents a basic design specification vector, including material and resolution; Puser represents a user creative description vector, including style and color; and α∈[0.6, 0.8] is a specification weight coefficient, which can be dynamically adjusted through industry standard design rules.

3. The system for clothing pattern personalization and customization based on the improved Stable Diffusion model according to claim 1 or 2, characterized in that, The core processing module further includes a CLIP alignment unit that maps the user-uploaded sketches and / or reference images to the same semantic space as the text prompt words, and generates visual consistency of the generated pattern through cosine similarity constraints. By cross-modal feature projection formula is represented, where F image represents an image feature vector; F text represents a text prompt word feature vector; ||F|| represents the Euclidean norm of the feature vector, and the similarity threshold > 0.75 triggers automatic optimization.

4. The system according to claim 3, wherein, In the TPS deformation of the dynamic adaptation engine, the energy minimization problem of thin plate spline interpolation satisfies wherein E represents an energy function of thin plate spline interpolation; represents the second derivative of the function f with respect to x to achieve control bending stiffness; represents a mixed second derivative to achieve control shear deformation.

5. The system of claim 1, wherein the system is based on an improved Stable Diffusion model. The core processing module further includes a Perlin noise unit that generates a natural transition density distribution map through a fractal noise algorithm, and controls the random density variation of pattern elements, and is represented by gradient field interpolation and an eight-degree noise superposition formula, which is: where D(x, y) represents the density distribution map, with a value range of [0, 1], 0 representing no elements and 1 representing the maximum density; i represents the index of the octave noise layer, i being any positive integer in 0 to k; noise(2 i x, 2 i y) represents the i-th layer Perlin noise value, with a frequency of 2 i , simulating natural texture details; H represents the Hurst index, with a value range of [0, 1]; 2 i(2-H) represents the amplitude attenuation factor, supporting the user to adjust the "complexity" parameter in real time, and the density coefficient D ∈ [0.1, 2.0].

6. A clothing pattern personalization customization method based on an improved Stable Diffusion model, characterized in that, The method includes the following steps: Multi-modal pattern generation: fuse the semantics of basic design specifications and user creative descriptions to generate a joint prompt word vector, combine user-uploaded sketches and / or reference images, and generate an initial clothing pattern through CLIP alignment; Dynamic adaptation and parameter control: based on clothing pattern key points, perform nonlinear geometric deformation on the initial pattern using the TPS algorithm, and then control the density of pattern elements using the Perlin noise algorithm to output the final pattern; Production file output: automatically generate vector files and process parameter packages that meet industrial standards to interface with production equipment.

7. The garment pattern personalization customization method according to claim 6, wherein, In the multi-modal pattern generation step, the joint prompt word vector generation formula is: f used＝ αP base +(1-α)P user , CLIP alignment generates pattern visual consistency by cosine similarity constraint, the formula is Similarity threshold > 0.75 triggers automatic optimization.

8. The garment pattern personalization customization method according to claim 6, wherein, In the dynamic adaptation and parameter control step, the TPS deformation is based on the thin plate spline interpolation energy minimization formula: Nonlinear geometric deformation is performed, and Perlin noise is superimposed by gradient field interpolation and octaves noise superposition formula: The density of the pattern elements is controlled, and the user can adjust the density coefficient D ∈ [0.1, 2.0].

Images