Research map / Image generation / Autoregressive and latent generation
TexGarment: Consistent Garment UV Texture Generation via Efficient 3D Structure-Guided Diffusion Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces TexGarment, a single-pass PixArt-α-based diffusion transformer that generates 3D-consistent garment UV textures by conditioning on a UV position map plus point-cloud structure latent.
Problem
Generating high-quality, diverse, and 3D-consistent garment textures in UV space, where 2D-to-3D projection methods suffer from view inconsistency and 3D-data-trained methods lack diversity and generalization; iterative methods are also slow.
Approach
Adapt a pre-trained text-to-image diffusion transformer (PixArt-α) to UV texture generation by adding a 3D structure latent, built by cross-attending UV position-map latents to point-cloud embeddings, and injecting it via an extra cross-attention branch alongside the text cross-attention.
Key results
- TexGarment reaches FID 96.97 and KID 0.63 vs. 147.40/2.86 for Paint3D, with CLIP score 32.37 and ~4 s runtime on a single V100.
- User study (30 users, 50 meshes) gives image quality 4.83 and text fidelity 4.92 for TexGarment vs. 4.01 and 4.17 for Paint3D.
- Ablation shows the combined UV position map plus point cloud guidance (FID 96.97) beats point-cloud-only (187.63), UV-position-only (174.29) and mask-only (210.34) guidance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add combined 2D UV layout and global 3D structure guidance | TexGarment · Garment texture synthesis | FID: 147.40 -> 96.97 | Helped |
| add combined 2D UV layout and global 3D structure guidance | TexGarment · Garment texture synthesis | Runtime: ∼220s -> ∼4s | Helped |
| add combined 2D UV layout and global 3D structure guidance | TexGarment · Garment texture synthesis | CLIP score: 28.72 -> 32.37 | Helped |
| add combined 2D UV layout and global 3D structure guidance | TexGarment · User study | Image quality: 4.01 -> 4.83 | Helped |
| add combined 2D UV layout and global 3D structure guidance | TexGarment · User study | Text fidelity: 4.17 -> 4.92 | Helped |
| add 2D UV layout information to point cloud guidance (ablation) | TexGarment · Garment texture synthesis | FID: 187.63 -> 96.97 | Helped |
| add global 3D information to UV position map guidance (ablation) | TexGarment · Garment texture synthesis | FID: 174.29 -> 96.97 | Helped |
| add 2D UV layout and global 3D structure information to mask-only guidance (ablation) | TexGarment · Garment texture synthesis | FID: 210.34 -> 96.97 | Helped |
Limitations
- Does not support image-conditioned generation (only text prompts).
- Trained on a private internal dataset of about 8k VRoid garments; evaluation uses a small 0.1k test split.
- Evaluation relies on FID/KID/CLIP plus a 30-user study; baseline TexPainter only compared qualitatively due to runtime.
Experimental setup
Datasets: VRoid Hub garment models (internal 8k set)
Benchmarks: internal garment test set (0.1k garments)
Metrics: FID KID CLIP score runtime user study (1-5 image quality and text fidelity)
Baselines: TEXTure Text2Tex Paint3D TexPainter
Builds on: PixArt-α DiT classifier-free guidance Paint3D
Closest papers by meaning
- DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.86)
- Reconstruction of Manipulated Garment with Guided Deformation Prior (NeurIPS 2024 · similarity 0.86)
- TexOct: Generating Textures of 3D Models with Octree-based Diffusion (CVPR 2024 · similarity 0.86)
- FashionR2R: Texture-preserving Rendered-to-Real Image Translation with Diffusion Models (NeurIPS 2024 · similarity 0.86)
- SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction (CVPR 2024 · similarity 0.85)
- Text-Guided 3D Face Synthesis - From Generation to Editing (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.