Over the past two years, AIGC has become increasingly tangible in fashion. Designers can generate visual inspiration from text, brands can produce marketing assets more quickly, and e-commerce teams are exploring digital models, virtual try-on, and automated content creation.
These developments are highly visible because they appear directly on screen and reach consumers immediately. But a garment must ultimately be made. A compelling image still needs to be translated into patterns, seams, fabric, and construction. How fabric pieces are cut, which edges are sewn together, and how they form a fitted 3D shape are all central to whether a design can move from concept into production.
There remains a significant gap between generating an image of a garment and generating a garment that can be edited, simulated, and developed further.
Style3D Research’s GarmageNet addresses this gap. Published in ACM Transactions on Graphics and presented at SIGGRAPH Asia 2025, the research was conducted jointly by Style3D, Zhejiang Sci-Tech University, Zhejiang University’s State Key Laboratory of CAD&CG, and Shanghai Jiao Tong University. Rather than focusing only on garment appearance, the work addresses a more difficult stage of fashion automation: the connection between patterns and 3D garments.
Why pattern generation is harder than image generation
A 3D garment is made from multiple 2D pattern pieces. Each piece has its own shape and dimensions, while the pieces are connected through specific sewing relationships. Once arranged around a body and simulated physically, they create the drape, wrinkles, and layered forms associated with real clothing.
Existing approaches generally fall into two categories. One begins with patterns, which produces clear structures that can be developed further, but may not capture detailed 3D drape in advance. The other generates 3D geometry directly, which can produce an appearance closer to a target design but may lack editable patterns and complete sewing structures. Without both structure and appearance, a result is difficult to carry into an established apparel workflow.
GarmageNet seeks to retain both 2D pattern structure and 3D garment form within one generative framework, then bring the generated result into physical simulation.
Training AI on garments closer to real production
Data is often the first challenge in garment-generation research.
Some existing datasets originate from body scans and preserve real wrinkles, but do not include corresponding production patterns. Others are generated procedurally and offer scale, yet may not reflect the multi-layer structures, complex silhouettes, local seams, and dense wrinkles commonly found in real production.
To address this, the team built GarmageSet, a dataset of 14,801 professionally created garments. It covers five main categories—tops, trousers, skirts, dresses, and outerwear—as well as sportswear, underwear, sleepwear, qipaos, and other styles. Each garment includes 2D patterns, 3D form, and detailed annotations at the level of pattern pieces, construction lines, and style features.
In a comparison reported in the paper, GarmageSet garments averaged 13.59 pattern pieces and 46.01 sewing relationships, compared with 10.82 and 30.26 respectively in GarmentCodeData. GarmageSet also uses substantially higher-resolution 3D meshes, enabling finer representation of wrinkles and structural details.
The dataset was assembled through a workflow close to real garment development: reusable components were organized from real patterns; professional pattern makers and designers classified and annotated them; and 3D garment specialists combined, modified, and completed the final garments. The resulting assets are intended to remain compatible with established garment-production workflows.
From design concepts to digital garments
GarmageNet introduces a unified representation called Garmage to connect 2D patterns and 3D garments. Each pattern piece is encoded as a structured geometry image: the alpha channel records its 2D outline, while color channels capture its 3D position. This allows planar structure and 3D form to be represented together.
The system then uses a variational autoencoder to compress pattern pieces into latent representations and a diffusion Transformer to learn how complete garments are generated. Text descriptions, sketches, product images, original patterns, and point clouds can all serve as inputs.
Generating pattern pieces alone is not sufficient. The system must also determine which edges should be sewn together. GarmageJigsaw samples along pattern boundaries, identifies points involved in sewing, and predicts corresponding point pairs. The system then extracts vector patterns, establishes sewing constraints, generates triangular meshes, and uses the 3D information encoded in Garmage to initialize physical simulation. This is how 2D pattern pieces become a 3D digital sample with simulated drape and wrinkles.
The workflow spans pattern generation, sewing reconstruction, 3D initialization, and physical simulation. Its aim is not only to produce a garment-like shape, but also to generate an asset that can be edited and used within existing garment-development tools.
What the experiments show
The paper evaluates GarmageNet across pattern quality, sewing reconstruction, 3D form, and generation efficiency.
In sewing reconstruction, GarmageJigsaw achieved 99.16% precision and 97.13% recall for sewing-point identification. In simulation-initialization tests on 150 complex patterns, GarmageNet achieved a 91.41% success rate, compared with 59.38% for rigid initialization and 93.75% for an optimization-based method.
In comparisons with Omage and Surf-D, the paper reports improvements across multiple metrics for geometric quality, generation diversity, and inference speed. The reported inference time for a single garment was approximately eight seconds, compared with 120 seconds and 25.7 seconds for the two baselines under the paper’s evaluation conditions.
The research team also invited 20 fashion designers, pattern makers, and 3D garment artists to evaluate results. In text-guided generation, participants preferred GarmageNet outputs in 85% of garment-aesthetics comparisons and 90.42% of pattern-aesthetics comparisons. In sketch-guided generation, preference for pattern aesthetics reached 97.66%.
These results should be interpreted within the paper’s datasets, baselines, and test conditions. They indicate that the approach can balance structural completeness, 3D results, and efficiency. They do not mean every garment category can already be generated for production without human review. The paper also identifies current limitations, including training data that is limited to a standard size-S avatar in an A-pose, as well as opportunities to improve support for internal structures such as pockets.
Where structured patterns and digital samples can be used
The paper demonstrates four direct application areas.
First, design generation. Text, sketches, and product images can be used to generate garment assets with pattern structure and 3D form, supporting early design exploration and pre-sampling review.
Second, digital sampling from 2D patterns. A flat DXF pattern without sewing relationships or 3D form can be completed into a simulation-ready digital sample. This can help teams see fit and structure earlier in the development process.
Third, pattern reconstruction from scans. Structured patterns can be recovered from point clouds created through scanning or multi-view reconstruction. This reconnects a 3D exterior shape with the pattern information required for garment development.
Fourth, progressive editing. Users can add sleeves, change a collar, or modify selected components while preserving the unedited parts of the garment. This is closer to the iterative way designers work than regenerating an entire garment from scratch.
Together, these applications connect design concepts, patterns, digital samples, and production communication. They can support design reviews, virtual sampling, pattern validation, product visualization, and pre-production collaboration, while also providing more complete digital assets for fabric simulation, dynamic presentation, and virtual try-on.
Advancing fashion AIGC toward production workflows
Fashion already has many tools that generate attractive images. Production workflows, however, require systems that can understand pattern pieces, sewing relationships, and 3D garment form.
GarmageNet offers one promising direction. Starting with data closer to real garment-development workflows, it uses a diffusion model to generate pattern structures and 3D initialization, then combines sewing reconstruction with physical simulation to create digital samples that can be developed further.
The work still has open questions around body-size coverage, complex construction, and industrial validation. Yet it advances the conversation around fashion AIGC from visual generation toward a deeper understanding of how garments are structured, assembled, and worn.
Open research for a growing community
GarmageNet was published in 2025 and presented at SIGGRAPH Asia 2025. The team has released the project page, code, and GarmageSet dataset publicly, enabling researchers and practitioners to study and build upon the work.
This is an example of how industry and academia can work together: Style3D contributes experience from digital fashion workflows and professionally annotated garment data, while collaborators from Zhejiang University, Shanghai Jiao Tong University, and Zhejiang Sci-Tech University contribute research expertise in computer graphics and AI.
For more information, visit the GarmageNet project page.
