DressCode is a framework for generating 3D garments from text descriptions, targeting fashion designers, virtual try-on applications, and digital human creation. It democratizes garment design by enabling natural language interaction to produce CG-friendly sewing patterns and PBR textures.

How It Works

DressCode employs a two-stage approach. First, SewingGPT, a GPT-based architecture, generates sewing patterns by integrating cross-attention with text-conditioned embeddings. This autoregressive method allows for detailed, text-guided pattern creation. Second, a fine-tuned Stable Diffusion model generates tile-based Physically-based Rendering (PBR) textures for the generated patterns, enhancing realism.

Quick Start & Requirements

• Install via conda env create -f environment.yaml and conda activate DressCode.
• Requires Python 3.x, Conda, and specific packages for simulation (see instructions).
• Download Stable Diffusion 2-1 base model and pretrained SewingGPT and PBR models.
• Configuration involves updating system.json with local paths for datasets, models, and rendering tools (Blender, Maya).
• Simulation is Windows-only.
• Project Page: https://github.com/IHe-KaiI/DressCode

Highlighted Details

• Generates both sewing patterns and PBR textures from text prompts.
• Supports pattern completion and texture editing via a Gradio UI.
• Offers interactive garment generation using ChatGPT integration.
• Can simulate and render results using Blender (Windows only) and export meshes for external software like Marvelous Designer.

Maintenance & Community

The project is associated with SIGGRAPH 2024 and lists multiple authors from academic institutions. No specific community channels (Discord, Slack) or roadmap are mentioned in the README.

Licensing & Compatibility

The repository does not explicitly state a license. The project is built upon NeuralTailor and Sewformer, whose licenses should be considered for compatibility.

Limitations & Caveats

Simulation and rendering capabilities are limited to Windows. Texture editing currently supports only one garment at a time. The dataset is based on external sources, and specific details about their licensing are not provided.