DressCode - A 3D clothing generation framework developed by ShanghaiTech University
DressCode is a 3D clothing generation framework jointly developed by ShanghaiTech University, the University of Pennsylvania, and Deemos Technologies. It allows users to automatically generate 3D clothing models of various styles and materials based on text descriptions.
What is DressCode?
DressCode is a 3D clothing generation framework jointly developed by ShanghaiTech University, the University of Pennsylvania, and Deemos Technologies. It allows users to automatically generate 3D clothing models of various styles and materials based on text descriptions. Built on the SewingGPT core module, DressCode understands text prompts and converts them into detailed cutting patterns. Combined with physically based rendering techniques, it creates realistic clothing effects.
DressCode's main functions
- Text-driven clothing generationThe user inputs a text description, and the system automatically generates a corresponding 3D clothing model.
- Material and texture simulationGenerate different materials, such as silk and lace, based on text prompts, and simulate realistic lighting effects.
- Semantic understanding and pattern generationThe SewingGPT module parses text and generates a sequence of cropped pattern tokens.
- Physically based renderingAdvanced fabric dynamics algorithm to simulate the drape and dynamic effects of real clothing.
The technical principle of DressCode
- Natural Language Processing (NLP)DressCode uses advanced NLP technology to parse and understand user text input, capturing key features and style requirements in clothing descriptions.
- Sequence generation modelSewingGPT, as the core component, adopts a Transformer-based decoder architecture to transform text descriptions into serialized representations (token sequences) of clothing cutting patterns.
- Quantization and dequantizationThe parameters of the clothing pattern are converted into tokens through a quantization process, and then dequantized after the generation process to reconstruct the pattern in 3D space.
- Conditional generationBy utilizing textual conditional embedding and cross-attention mechanisms, SewingGPT can generate clothing patterns that match the description based on textual prompts.
- Physically Based Rendering (PBR)DressCode uses PBR technology to generate realistic textures and material effects for clothing, simulating the reflection and refraction characteristics of different fabrics under light.
- Fabric dynamics simulationIt integrates advanced fabric simulation algorithms to ensure that clothing exhibits realistic physical behavior in a virtual environment, such as wrinkles and swaying.
DressCode's project address
-
Project official website:https://ailb-web.ing.unimore.it/dress-code/dataset
-
GitHubstorehouse:https://github.com/aimagelab/dress-code
-
arXivTechnical Papers:https://arxiv.org/abs/2204.08532
Application scenarios of DressCode
- Fashion DesignDesigners can quickly generate clothing prototypes using text descriptions, accelerating the design process and enabling rapid iteration.
- Virtual try-onE-commerce platforms and fashion brands can use DressCode to offer virtual try-on services, allowing consumers to experience how clothing looks online.
- Game developmentGame designers can use DressCode to quickly generate costumes for game characters, enriching the in-game appearance options.
- Film and television productionIn the costume preparation stage of movies and TV series, DressCode can help designers quickly generate costume concept images based on script descriptions.