**DressCode: Autoregressively Sewing and Generating Garments from Text Guidance** **Authors:** KAI HE, KAIXIN YAO, QIXUAN ZHANG, JINGYI YU, LINGJIE LIU, LAN XU **Abstract:** Apparel plays a significant role in human appearance, making garment digitalization crucial for digital human creation. Recent advances in 3D content creation have been pivotal, but garment generation from text guidance remains nascent. This paper introduces DressCode, a text-driven 3D garment generation framework that aims to democratize design for novices and offer immense potential in fashion design, virtual try-on, and digital human creation. The framework includes SewingGPT, a GPT-based architecture that integrates cross-attention with text-conditioned embedding to generate sewing patterns, and a pre-trained Stable Diffusion model tailored to generate tile-based Physically-based Rendering (PBR) textures for the garments. By leveraging a large language model, DressCode generates CG-friendly garments through natural language interaction, facilitating pattern completion and texture editing. Comprehensive evaluations and comparisons with state-of-the-art methods demonstrate superior quality and alignment with input prompts. User studies further validate the high-quality rendering results, highlighting the practical utility and potential in production settings. **Contributions:** - A first text-driven garment generation pipeline with high-quality sewing patterns and PBR textures. - A novel generative paradigm for sewing patterns as a sequence of tokens, achieving high-quality autoregressive generation via text guidance. - A tailored diffusion model for vivid texture generation of garments from text prompts, showcasing interaction-friendly applications for garment generation, completion, and editing. **Methods:** - **SewingGPT:** A GPT-based autoregressive model for sewing pattern generation using text prompts. It converts sewing patterns into token sequences and generates them autoregressively. - **PBR Texture Generation:** A pre-trained Stable Diffusion model is fine-tuned to generate tile-based PBR textures from text prompts, ensuring high-quality rendering. **Applications:** - **Garment Generation:** Users can generate diverse sewing patterns and corresponding PBR textures through natural language interaction. - **Pattern Completion:** The method can complete entire sewing patterns based on partial information and text prompts. - **Texture Editing:** Structured UV mappings enable efficient texture editing at specific locations. **Experiments:** - **Qualitative and Quantitative Comparisons:** DressCode outperforms state-of-the-art methods in generating accurate and detailed sewing patterns and PBR textures. - **Ablation Study:** The effectiveness of the triple embedding used in SewingGPT is evaluated, showing that positional and parameter embeddings significantly improve the quality of generated patterns. - **User Study:** A comprehensive user study demonstrates the higher user preference for DressCode's generated garments compared to competing approaches. **Limitations and Future Work:** - The method struggles with prompts outside the domain of the training dataset and**DressCode: Autoregressively Sewing and Generating Garments from Text Guidance** **Authors:** KAI HE, KAIXIN YAO, QIXUAN ZHANG, JINGYI YU, LINGJIE LIU, LAN XU **Abstract:** Apparel plays a significant role in human appearance, making garment digitalization crucial for digital human creation. Recent advances in 3D content creation have been pivotal, but garment generation from text guidance remains nascent. This paper introduces DressCode, a text-driven 3D garment generation framework that aims to democratize design for novices and offer immense potential in fashion design, virtual try-on, and digital human creation. The framework includes SewingGPT, a GPT-based architecture that integrates cross-attention with text-conditioned embedding to generate sewing patterns, and a pre-trained Stable Diffusion model tailored to generate tile-based Physically-based Rendering (PBR) textures for the garments. By leveraging a large language model, DressCode generates CG-friendly garments through natural language interaction, facilitating pattern completion and texture editing. Comprehensive evaluations and comparisons with state-of-the-art methods demonstrate superior quality and alignment with input prompts. User studies further validate the high-quality rendering results, highlighting the practical utility and potential in production settings. **Contributions:** - A first text-driven garment generation pipeline with high-quality sewing patterns and PBR textures. - A novel generative paradigm for sewing patterns as a sequence of tokens, achieving high-quality autoregressive generation via text guidance. - A tailored diffusion model for vivid texture generation of garments from text prompts, showcasing interaction-friendly applications for garment generation, completion, and editing. **Methods:** - **SewingGPT:** A GPT-based autoregressive model for sewing pattern generation using text prompts. It converts sewing patterns into token sequences and generates them autoregressively. - **PBR Texture Generation:** A pre-trained Stable Diffusion model is fine-tuned to generate tile-based PBR textures from text prompts, ensuring high-quality rendering. **Applications:** - **Garment Generation:** Users can generate diverse sewing patterns and corresponding PBR textures through natural language interaction. - **Pattern Completion:** The method can complete entire sewing patterns based on partial information and text prompts. - **Texture Editing:** Structured UV mappings enable efficient texture editing at specific locations. **Experiments:** - **Qualitative and Quantitative Comparisons:** DressCode outperforms state-of-the-art methods in generating accurate and detailed sewing patterns and PBR textures. - **Ablation Study:** The effectiveness of the triple embedding used in SewingGPT is evaluated, showing that positional and parameter embeddings significantly improve the quality of generated patterns. - **User Study:** A comprehensive user study demonstrates the higher user preference for DressCode's generated garments compared to competing approaches. **Limitations and Future Work:** - The method struggles with prompts outside the domain of the training dataset and