Full-forming knitting design generation method based on low-rank adaptation and information constraint
The pre-trained model is fine-tuned and trained through the LoRA model, and combined with process parameter information, a design drawing that meets the requirements of the full molding process is generated, which solves the problem that the existing AI design generation system cannot be actually produced and achieves efficient design generation.
Patent Information
- Application Number
- CN202510505230.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-08-12
AI Technical Summary
Although the design drawings generated by the existing AI design generation system in the field of full-form knitting visually meet the requirements, they often violate the basic process principles of full-form knitting, resulting in the design of the clothing being unable to be actually produced.
The fully formed knitted design generation method based on low rank adaptation and information constraints is adopted. The pre-trained model is fine-tuned and trained through the LoRA model, and combined with process parameter information, a design drawing that meets the requirements of the full forming process is generated.
It realizes that the AI system can learn and follow the information constraints of fully molded knitted, and directly generates design solutions that are both beautiful and practically produced, reducing the designer's modification workload and improving design efficiency.
Smart Images
Figure CN120470893A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of fine-tuning pre-training models, and in particular to the design of fully fashioned knitwear. Background Art
[0002] In the field of knitwear design, wholly fashioned fabrics are an advanced, sewing-free knitting technology. Currently, designers must draw on years of experience and consider the various process parameters of the flat knitting machine. For example, when designing a turtleneck, the needle's retraction and expansion patterns must be considered to ensure a smooth neckline. When designing an armhole, the knitting machine's loop-forming mechanism must be considered to avoid unknittable structures.
[0003] At the same time, with the development of AI technology, AI-based design generation systems have emerged, such as those that use diffusion models to directly generate knitting designs. However, in the field of fully fashioned knitting, design must not only consider aesthetics but also meet certain process information constraints. Specifically, the design must strictly adhere to process limitations, such as the motion patterns of the knitting machine's loop-forming mechanism and the needle-retraction and -release rules. Existing knitting design generation technologies suffer from a fundamental problem: while the generated design drawings may meet the visual requirements, they often violate the basic process principles of fully fashioned knitting, resulting in garments that are unmanufactured. For example, neckline designs may narrow too quickly, exceeding the knitting machine's narrowing capacity; armhole designs may result in curves that do not conform to the knitting machine's loop-forming mechanism's motion patterns. This results in designers spending a significant amount of time revising the generated designs, severely impacting design efficiency. Summary of the Invention
[0004] The present invention proposes a fully fashioned knitwear design generation method based on low-rank adaptation and information constraints, which solves the problem that existing AI design generation systems can create styles with good visual effects, but often do not meet the basic detail requirements and process characteristics of these knitwear, resulting in the designed garments being unable to be actually produced.
[0005] The method for generating a fully fashioned knitted design based on low-rank adaptation and information constraints of the present invention comprises the following steps: Step S1: Obtain design requirement text and process parameter information text; Step S2: obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Step S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
[0006] Furthermore, a preferred embodiment is provided, wherein step S3 includes: The fine-tuned pre-trained model includes a CLIP model; The text editor of the CLIP model based on the fine-tuned pre-trained model encodes the design requirement text and process parameter information text into feature vectors; Noise initialization based on eigenvector: Generate a random noise matrix with fixed pixels as the initial image; Iterative denoising is performed based on the initial image to generate a design that meets the requirements of the full molding process; In each round of iterative denoising process: Use the UNet model to predict the noise residue of the current image, and gradually generate image details based on the feature vector and the noise residue; The adaptation layer of the LoRA model dynamically adjusts the generation process to ensure that the generated image meets the process information constraints of the fully fashioned knitted design.
[0007] Furthermore, a preferred embodiment is provided, in step S2, the method of fine-tuning the pre-trained model using the LoRA model is as follows: Step S2.1: Obtain a fully fashioned knitwear design dataset in the form of text-image feature pairs; each piece of data in the fully fashioned knitwear design dataset contains clothing style information, clothing structure information, tissue structure information, and process parameter information; Step S2.2: Convert the data of the fully fashioned knitted design dataset into feature vectors that can be learned by the pre-training model; Step S2.3: Based on the learnable feature vector of the pre-trained model, the LoRA model is used for fine-tuning training to obtain the parameters of the LoRA model containing process parameter information constraints, and the parameters of the LoRA model containing process parameter information constraints are combined with the pre-trained model to obtain a fine-tuned pre-trained model.
[0008] Furthermore, a preferred embodiment is provided, in step S2.3, based on the learnable feature vectors of the pre-trained model, the LoRA model is used for fine-tuning training as follows: The original weight matrix parameters of the pre-trained model remain fixed; Insert the low-rank matrix of the LoRA model into the self-attention layer of the pre-trained model; Low-rank approximation is achieved through matrix decomposition, and the calculated results are updated through forward propagation to obtain the parameters of the LoRA model containing process parameter information constraints.
[0009] Furthermore, a preferred embodiment is provided, wherein the pre-training model is a Stable Diffusion model.
[0010] Furthermore, a preferred embodiment is provided. In the method of fine-tuning the pre-trained model using the LoRA model, the training configuration of the LoRA model is as follows: Matrix rank: 32; Learning rate: 1e-4; Batch size: 4; Training epochs: 50; Optimizer: AdamW8bit.
[0011] The present invention also proposes a fully fashioned knitted design generation device based on low-rank adaptation and information constraints, which includes the following modules: Module S1: Obtain design requirement text and process parameter information text; Module S2: Obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Module S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
[0012] The present invention also proposes a computer device comprising: a processor and a memory, wherein the memory is used to store executable instructions of the processor, and the processor is configured to execute any one of the above-mentioned methods for generating fully formed knitted designs based on low-rank adaptation and information constraints by executing the executable instructions.
[0013] The present invention also proposes a computer storage medium, in which a computer program is stored. When the computer program is run, any one of the above-mentioned methods for generating fully fashioned knitted designs based on low-rank adaptation and information constraints is executed.
[0014] The present invention also proposes a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any one of the above-mentioned methods for generating fully fashioned knitted designs based on low-rank adaptation and information constraints.
[0015] The present invention has the following beneficial effects: 1. The method for generating fully fashioned knitwear designs based on low-rank adaptation and information constraints described in this invention uses LoRA (Low-Rank Adaptation) to fine-tune a pre-trained model, enabling an AI (Artificial Intelligence) system to learn and follow the information constraints of fully fashioned knitwear, directly generating designs that are both aesthetically pleasing and practically manufacturable.
[0016] 2. The method for generating fully fashioned knitwear designs based on low-rank adaptation and information constraints described in the present invention converts the (process) information constraints of fully fashioned knitwear into learnable features for the model, and implements targeted fine-tuning of the pre-trained model through the LoRA (Low-Rank Adaptation) technology. This method enables the model to naturally follow process laws while maintaining the innovativeness of the design, generating a design solution that can be actually produced.
[0017] 3. The method for generating fully fashioned knitwear designs based on low-rank adaptation and information constraints described in this invention directly addresses the problem of existing AI (artificial intelligence) design systems ignoring (process) information constraints. By learning from practical design samples, the system can generate design solutions that meet both aesthetic requirements and process constraints, greatly reducing the designer's modification workload. This not only improves design efficiency but also provides more reliable technical support for design innovation.
[0018] The method for generating fully-fashioned knitting designs based on low-rank adaptation and information constraints described in the present invention is suitable for generating fully-fashioned knitting designs. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0020] Figure 1 A schematic flow chart of a method for generating fully-fashioned knitwear designs based on low-rank adaptation and information constraints in one embodiment of the present invention; Figure 2 In one embodiment of the present invention, a fully fashioned knitted design dataset comprises a schematic diagram of data; Figure 3 Schematic diagram of text-image feature pairs in a fully fashioned knitwear design dataset in one embodiment of the present invention; Figure 4 A schematic diagram of a process for converting data from a fully fashioned knitting design dataset into a feature vector that can be learned by a pre-training model in one embodiment of the present invention; Figure 5 A schematic diagram of a process for fine-tuning training using the LoRA model in one embodiment of the present invention; Figure 6 In one embodiment of the present invention, the design drawing corresponding to the LoRA raw image numbered 1 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 7 In one embodiment of the present invention, the design drawing corresponding to the non-LoRA raw image numbered 1 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 8 In one embodiment of the present invention, the design drawing corresponding to the LoRA raw image numbered 2 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 9 In one embodiment of the present invention, the design drawing corresponding to the non-LoRA raw image numbered 2 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 10 In one embodiment of the present invention, the design drawing corresponding to the LoRA raw image numbered 3 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 11 In one embodiment of the present invention, the design drawing corresponding to the non-LoRA raw image numbered 3 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 12 In one embodiment of the present invention, the design drawing corresponding to the LoRA raw image numbered 4 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 13 In one embodiment of the present invention, the design drawing corresponding to the non-LoRA raw image numbered 4 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 14 In one embodiment of the present invention, the design drawing corresponding to the LoRA raw image numbered 5 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images); Figure 15 In one embodiment of the present invention, the design drawing corresponding to the non-LoRA raw image numbered 5 in Table 1 (i.e., a comparison table of LoRA raw images and non-LoRA raw images) is shown. DETAILED DESCRIPTION
[0021] In order to make the technical solutions and advantages of the present invention more clearly described, the specific embodiments of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various embodiments described below are only part of the preferred embodiments of the present invention, rather than all implementation plans; the various embodiments described below are intended to explain the present invention and cannot be understood as limiting the present invention; the reasonable combination of the technical features defined in the various embodiments of the present invention, as well as all other implementation plans obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work, all fall within the scope of protection of the present invention.
[0022] In one embodiment, a method for generating a fully fashioned knitted design based on low-rank adaptation and information constraints is provided, the method comprising the following steps: Step S1: Obtain design requirement text and process parameter information text; Step S2: obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Step S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
[0023] In this embodiment, the design requirement text is used to describe the clothing style information, clothing structure information, and tissue structure information of a wholly fashioned knitted design, such as "wholly fashioned turtleneck sweater."
[0024] In this embodiment, the process parameter information text is used to describe the process parameter information of the fully fashioned knitting design, such as the flat knitting machine model.
[0025] In this implementation, the LoRA model refers to low-rank adaptation.
[0026] It should be noted that compared to solutions that directly use rule constraints or rely entirely on parameterized design, the LoRA (Low-Rank Adaptation) fine-tuning method maintains design flexibility while ensuring process feasibility: The solution of directly using rule constraints is too rigid and difficult to adapt to the diversity of design; Solutions that rely entirely on parametric design limit the space for innovation.
[0027] In addition, in one embodiment, step S3 includes: The fine-tuned pre-trained model includes a CLIP model; The text editor of the CLIP model based on the fine-tuned pre-trained model encodes the design requirement text and process parameter information text into feature vectors; Noise initialization based on eigenvector: Generate a random noise matrix with fixed pixels as the initial image; Iterative denoising is performed based on the initial image to generate a design that meets the requirements of the full molding process; In each round of iterative denoising process: Use the UNet model to predict the noise residue of the current image, and gradually generate image details based on the feature vector and the noise residue; The adaptation layer of the LoRA model dynamically adjusts the generation process to ensure that the generated image meets the process information constraints of the fully fashioned knitted design.
[0028] In this embodiment, the CLIP model refers to Contrastive Language-Image Pre-training. The CLIP model includes a text editor and an image editor.
[0029] In this embodiment, the number of iterative denoising rounds is 30.
[0030] In this implementation, the "LoRA model adaptation layer" introduces trainable parameters into the pre-trained model's self-attention layer via low-rank matrix decomposition (A and B). Its purpose is to dynamically incorporate process constraints related to fully fashioned knitwear designs (such as armhole curves and weave transitions) by adjusting the intermediate feature representations during the generation process, ensuring that the generated images meet specific domain requirements.
[0031] In this implementation, the LoRA model forms an adaptation layer by inserting low-rank matrices (A and B) into the self-attention layer of a pre-trained model (such as Stable Diffusion). The original weights of the pre-trained model (such as the Transformer weight matrix W) remain fixed. The adaptation layer is a newly added trainable part of the LoRA model and is independent of the pre-trained model structure.
[0032] In this implementation, when using LoRA fine-tuning, the weight matrices (A and B) of the LoRA model are modified. The original weights (W) of the pre-trained model remain frozen during fine-tuning, and only the low-rank matrices (A and B) are used to adapt the model output.
[0033] In this embodiment, after fine-tuning training is completed, the parameters (A and B) of the LoRA model are used in combination with the pre-trained model, but the pre-trained model itself is not modified.
[0034] The fine-tuned pre-trained model refers to a pre-trained model that combines the parameters (A and B) of the LoRA model. The parameters (A and B) of the LoRA model contain process parameter information constraints (also known as process information constraints or process feature rules).
[0035] In addition, in one embodiment, in step S2, the method for fine-tuning the pre-trained model using the LoRA model is as follows: Step S2.1: Obtain a fully fashioned knitwear design dataset in the form of text-image feature pairs; each piece of data in the fully fashioned knitwear design dataset contains clothing style information, clothing structure information, tissue structure information, and process parameter information; Step S2.2: Convert the data of the fully fashioned knitted design dataset into feature vectors that can be learned by the pre-training model; Step S2.3: Based on the learnable feature vector of the pre-trained model, the LoRA model is used for fine-tuning training to obtain the parameters of the LoRA model containing process parameter information constraints, and the parameters of the LoRA model containing process parameter information constraints are combined with the pre-trained model to obtain a fine-tuned pre-trained model.
[0036] In this implementation, the model automatically learns and identifies which designs are most likely to be implemented in production by studying historically successful fully-fashioned knitwear examples. By learning from the patterns of successful designs, the model can generate designs that meet process requirements and have high production feasibility, significantly improving the match between design and actual production.
[0037] In this implementation, the fully fashioned knitwear design dataset includes 320 design data (or samples). These are all fully fashioned knitwear designs that have been verified in actual production and include various combinations of collar shapes, sleeve shapes, and weave structures. These samples fully demonstrate the characteristics and limitations of fully fashioned knitwear, providing a reliable learning foundation for the model.
[0038] In this embodiment, the clothing style information includes style (pullover, cardigan, etc.), body (loose, fitted, etc.), sleeve length, length of clothing, etc.
[0039] In this embodiment, the clothing structure information includes collar type (high collar, round collar, etc.) and sleeve type (raglan sleeve, batwing sleeve, etc.).
[0040] In this embodiment, the tissue structure information includes neckline tissue, cuff tissue, hem tissue, etc.
[0041] In this embodiment, the process parameter information includes the flat knitting machine model, yarn specifications, etc.
[0042] In this implementation, LoRA (Low-Rank Adaptation) technology is used to fine-tune the model for the specific domain of fully-fashioned knitwear. During fine-tuning, the model learns detailed information constraints specific to fully-fashioned knitwear, such as the armhole shaping curve and the transition requirements of the weave structure. This knowledge (process information constraints) is encoded in the LoRA (Low-Rank Adaptation) weight matrix, enabling the model to automatically consider these process constraints when generating designs.
[0043] In this embodiment, by fine-tuning the pre-trained model using the LoRA model, the intelligent generation of fully-fledged knitted designs is achieved, ensuring that the generated designs meet both aesthetic requirements and process information constraints.
[0044] In addition, in one embodiment, the step S2.2: converting the data of the fully fashioned knitting design dataset into a feature vector that can be learned by the pre-training model is as follows: The pre-trained model includes a CLIP model; the CLIP model includes a text encoder and an image encoder; A text encoder is used to convert the text data in the fully fashioned knitted design dataset into feature vectors that can be learned by the pre-training model; An image encoder is used to convert image data in the fully fashioned knitted design dataset into feature vectors that can be learned by the pre-trained model.
[0045] In this implementation, the text encoder is a CLIP text encoder.
[0046] In this implementation, the image encoder is a RESNET image encoder.
[0047] In this embodiment, the learnable feature vectors obtained by the text encoder and the image encoder are fused in the feature fusion layer to obtain a fused learnable feature vector; The fused learnable feature vector is normalized by a normalization layer to obtain a normalized learnable feature vector.
[0048] In addition, in one embodiment, in step S2.3, based on the learnable feature vectors of the pre-trained model, the LoRA model is used for fine-tuning training as follows: The original weight matrix parameters of the pre-trained model remain fixed; Insert the low-rank matrix of the LoRA model into the self-attention layer of the pre-trained model; Low-rank approximation is achieved through matrix decomposition, and the calculated results are updated through forward propagation to obtain the parameters of the LoRA model containing process parameter information constraints.
[0049] In this embodiment, the pre-trained model includes a Transformer model, and the Transformer model includes a self-attention layer.
[0050] In this embodiment, the original weight matrix of the pre-trained model is W. The low-rank matrices are A and B.
[0051] In this embodiment, the knowledge of process parameter information constraints (or process feature rules) is encoded in the weight matrix of the LoRA model by introducing low-rank matrices (A and B), matrix decomposition, and forward propagation.
[0052] In addition, in one embodiment, the pre-trained model is a Stable Diffusion model.
[0053] In addition, in one embodiment, the version of the Stable Diffusion model is v2.1, ie, the Stable Diffusion v2.1 model.
[0054] In addition, in one embodiment, in the method of fine-tuning the pre-trained model using the LoRA model, the training configuration of the LoRA model is as follows: Matrix rank: 32; Learning rate: 1e-4; Batch size: 4; Training epochs: 50; Optimizer: AdamW8bit.
[0055] In this embodiment, AdamW8bit represents adaptive momentum estimation.
[0056] In addition, in one embodiment, in the method of fine-tuning the pre-trained model using the LoRA model, the inference generation configuration of the LoRA model is as follows Iterations: 35 Cue word relevance: 7.
[0057] In one embodiment, a fully fashioned knitted design generation apparatus based on low-rank adaptation and information constraints is provided, the apparatus comprising the following modules: Module S1: Obtain design requirement text and process parameter information text; Module S2: Obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Module S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
[0058] In one embodiment, a computer device is provided, comprising: a processor and a memory, wherein the memory is used to store executable instructions of the processor, and the processor is configured to execute any one of the above-mentioned methods for generating a fully formed knitted design based on low-rank adaptation and information constraints by executing the executable instructions.
[0059] In one embodiment, a computer storage medium is provided, wherein a computer program is stored in the storage medium. When the computer program is run, any one of the above-mentioned methods for generating a fully fashioned knitted design based on low-rank adaptation and information constraints is executed.
[0060] In one embodiment, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any one of the above-described methods for generating a fully fashioned knitted design based on low-rank adaptation and information constraints.
[0061] In one embodiment, a fully fashioned knitwear design generation system based on low-rank adaptation and information constraints is provided, the system comprising: a training module and an inference module; The training module includes a data acquisition module, a feature extraction module and a LoRA fine-tuning module; The data acquisition module is used to acquire a fully fashioned knitted design data set in the form of text-image feature pairs; each piece of data in the fully fashioned knitted design data set contains clothing style information, clothing structure information, tissue structure information, and process parameter information; The feature extraction module converts the data of the fully fashioned knitted design dataset into a feature vector that can be learned by the pre-training model; The LoRA fine-tuning module: based on the learnable feature vector of the pre-trained model, fine-tuning training is performed using the LoRA model to obtain parameters of the LoRA model containing process parameter information constraints, and the parameters of the LoRA model containing process parameter information constraints are combined with the pre-trained model to obtain a fine-tuned pre-trained model; The reasoning module includes a design requirement acquisition module, a model deployment module and a design generation module; The design requirement input: obtaining the design requirement text and process parameter information text; The model deployment module is configured to obtain a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; The design generation module uses a fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process according to the design requirement text and the process parameter information text.
[0062] In addition, in one embodiment, in order to verify the manufacturability of the method, a comparative experiment is conducted to demonstrate the advantages of the method in improving the manufacturability of the design.
[0063] The impact of LoRA model fine-tuning technology is examined through comparative experiments.
[0064] In the comparative experiment, 100 fully fashioned sweater design images were generated using and not using the LoRA model fine-tuning technology.
[0065] A qualitative comparison of the raw image quality is performed, and the comparison results are as follows: For the wholly-furnished concept features: The sweater designed using LoRA model fine-tuning technology has no obvious seams, but you can see that the sweater has been connected by needles at the seams to ensure the connection of the fabric, which can better demonstrate the fully molded characteristics.
[0066] The sweater in the design generated without the LoRA model fine-tuning technology has obvious seams.
[0067] For clothing styles and silhouettes: The designs generated using the LoRA model fine-tuning technology and the designs generated without the LoRA model fine-tuning technology both meet the style prompt words; However, in terms of silhouette generation, the design generated without using the LoRA model fine-tuning technology cannot understand the prompt word well; For the generation of clothing parts, since the shoulder and sleeve connection method of a fully fashioned sweater is different from that of an ordinary sweater, the design generated using LoRA model fine-tuning technology can better understand its characteristics and generate it.
[0068] For collars, both designs generated using LoRA model fine-tuning technology and designs generated without LoRA model fine-tuning technology can learn and generate collar features well.
[0069] As shown in Table 1, there is a comparison table of the design generated by using the LoRA model fine-tuning technology (referred to as LoRA raw images) and the design generated without using the LoRA model fine-tuning technology (referred to as non-LoRA raw images), referred to as the comparison table of LoRA raw images and non-LoRA raw images.
[0070] Table 1: Comparison of LoRA raw images and non-LoRA raw images
[0071] In the table, Figures 6 to 15 See the attached drawings of the specification.
[0072] It should be noted that there are significant differences between fully fashioned knitting and ordinary knitting in terms of process principles, equipment requirements, design constraints and production processes. A characteristic of conventional knitwear is that garment pieces must be sewn together using a sewing machine, resulting in variations in seams and thickness. This is due to the process of flat knitting, design, cutting, and sewing. During the knitting process, a piece with a relatively fixed stitch count is first produced. This allows for rapid design verification through sample cutting, and structural adjustments can be made through subsequent cutting. This results in lower requirements for knitting equipment, a certain tolerance for production errors, and a high degree of design freedom.
[0073] A hallmark of fully fashioned knitting is the seamless transition between garment pieces, achieved through the use of narrowing and widening needles. This is achieved through the three-dimensional, seamless knitting process. The knitting machine adjusts the number of needles (narrowing / widening) in real time during the knitting process to create complex curvatures, completing the three-dimensional structure of the garment pieces in a single process. This includes areas such as the armhole, neckline, and shoulder line. For example, the creation of a fully fashioned three-dimensional collar requires dynamic narrowing to create an arc-shaped curve during knitting, and the narrowing rate must match the knitting speed of the flat knitting machine's loop forming mechanism. Round collars require uniform narrowing.
[0074] Because the number of needles must be adjusted in real time during the knitting process, fully fashioned knitting equipment requires high standards. It must support dynamic needle expansion and contraction, multiple needle gauge switching, and coordinated yarn feeder motion to prevent incorrect expansion and contraction or yarn breakage, which could result in the entire garment being scrapped. This results in a low tolerance for error in the generated solutions. Furthermore, the integrated knitting process involves design, programming, knitting, and finishing (no cutting or sewing). A correct design is crucial for a smooth knitting process, as failure to do so would lead to resource depletion.
[0075] This research addresses the high cost of real-world trial knitting by proposing a solution: using virtual simulation combined with physical trial knitting to predict narrowing effects and generate fully fashioned knitting designs. Design generation for fully fashioned knitting requires deep integration of process constraints, and traditional AI generation methods (such as general diffusion models) lack domain-specific information, making them inadequate for actual production needs. This innovative solution addresses this issue through LoRA fine-tuning combined with process constraint encoding.
[0076] This embodiment provides a computer device or system, the hardware device of this part is a general model and is not shown in the form of a diagram. The system includes a processor and a memory, wherein the processor and the memory can be connected through a bus or other means. The memory is a non-transient computer-readable storage medium that can be used to store non-transient software programs, non-transient computer executable programs and modules, and corresponding program instructions / modules. The processor executes various functional applications and data processing of the processor by running the non-transient software programs, instructions and modules stored in the memory, so as to realize the data space entity resolution data quality enhancement method in the above method embodiment.
[0077] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, an intranet, a mobile communication network, and combinations thereof.
[0078] One or more modules are stored in the memory. When the processor executes, the method steps in the embodiment are executed. In this way, the purpose of the invention can be achieved through the method, device and process of the present invention. The specific details of the above-mentioned computer equipment can be understood by referring to the corresponding descriptions and effects in the embodiment, and will not be repeated here.
[0079] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD). The storage medium can also include a combination of the above-mentioned types of memory.
[0080] The above further describes the technical solution provided by the present invention in detail through several specific embodiments in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific embodiments described above are not intended to limit the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation methods and equivalent replacements based on the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A fully fashioned knitted design generation method based on low-rank adaptation and information constraints, characterized by: The method comprises the following steps: Step S1: Obtain design requirement text and process parameter information text; Step S2: obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Step S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
2. The method for generating fully fashioned knitted designs based on low-rank adaptation and information constraints according to claim 1, characterized in that: The step S3 comprises: The fine-tuned pre-trained model includes a CLIP model; The text editor of the CLIP model based on the fine-tuned pre-trained model encodes the design requirement text and process parameter information text into feature vectors; Noise initialization based on eigenvector: Generate a random noise matrix with fixed pixels as the initial image; Iterative denoising is performed based on the initial image to generate a design that meets the requirements of the full molding process; In each round of iterative denoising process: Use the UNet model to predict the noise residue of the current image, and gradually generate image details based on the feature vector and the noise residue; The adaptation layer of the LoRA model dynamically adjusts the generation process to ensure that the generated image meets the process information constraints of the fully fashioned knitted design.
3. The method for generating fully fashioned knitted designs based on low-rank adaptation and information constraints according to claim 1, characterized in that: In step S2, the method for fine-tuning the pre-trained model using the LoRA model is as follows: Step S2.1: Obtain a fully fashioned knitwear design dataset in the form of text-image feature pairs; each piece of data in the fully fashioned knitwear design dataset contains clothing style information, clothing structure information, tissue structure information, and process parameter information; Step S2.2: Convert the data of the fully fashioned knitted design dataset into feature vectors that can be learned by the pre-training model; Step S2.3: Based on the learnable feature vector of the pre-trained model, the LoRA model is used for fine-tuning training to obtain the parameters of the LoRA model containing process parameter information constraints, and the parameters of the LoRA model containing process parameter information constraints are combined with the pre-trained model to obtain a fine-tuned pre-trained model.
4. The method for generating fully fashioned knitted designs based on low-rank adaptation and information constraints according to claim 3, characterized in that: In step S2.3, based on the learnable feature vectors of the pre-trained model, the LoRA model is used for fine-tuning training as follows: The original weight matrix parameters of the pre-trained model remain fixed; Insert the low-rank matrix of the LoRA model into the self-attention layer of the pre-trained model; Low-rank approximation is achieved through matrix decomposition, and the calculated results are updated through forward propagation to obtain the parameters of the LoRA model containing process parameter information constraints.
5. The method for generating fully fashioned knitted designs based on low-rank adaptation and information constraints according to claim 3, characterized in that: The pre-training model is a Stable Diffusion model.
6. The method for generating fully fashioned knitted designs based on low-rank adaptation and information constraints according to claim 3, characterized in that: In the method of fine-tuning the pre-trained model using the LoRA model, the training configuration of the LoRA model is as follows: Matrix rank: 32; Learning rate: 1e-4; Batch size: 4; Training epochs: 50; Optimizer: AdamW8bit.
7. A fully fashioned knitting design generation device based on low-rank adaptation and information constraints, characterized in that: The device comprises the following modules: Module S1: Obtain design requirement text and process parameter information text; Module S2: Obtaining a fine-tuned pre-trained model; the fine-tuned pre-trained model is obtained by fine-tuning the pre-trained model using the LoRA model; the fine-tuned pre-trained model includes process parameter information constraints; Module S3: Use the fine-tuned pre-trained model to generate a design drawing that meets the requirements of the full molding process based on the design requirement text and process parameter information text.
8. A computer device comprising: A processor and a memory, characterized in that the memory is used to store executable instructions of the processor, and the processor is configured to execute the fully formed knitted design generation method based on low-rank adaptation and information constraints according to any one of claims 1 to 6 by executing the executable instructions.
9. A computer storage medium, characterized in that The storage medium stores a computer program, and when the computer program is run, the method for generating a fully fashioned knitted design based on low-rank adaptation and information constraints according to any one of claims 1 to 6 is executed.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the method for generating a fully fashioned knitted design based on low-rank adaptation and information constraints according to any one of claims 1 to 6 are implemented.