Research map / Image generation / Editing and style transfer
GenesisTex: Adapting Image Denoising Diffusion to Texture Space
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.
Proposes GenesisTex, a training-free text-to-texture method that runs DDIM-style denoising directly on per-view latent texture maps with cross-view style consistency and dynamic latent alignment, followed by inpainting and img2img refinement.
Problem
Generating high-quality, multi-view-consistent textures for given 3D meshes from text prompts, while avoiding the slow optimization of SDS and the view-order dependence and seam artifacts of inpainting-based methods.
Approach
Performs diffusion sampling in texture space by maintaining a latent texture map per viewpoint that is denoised concurrently via rendering and inverse rendering, with cross-view self-attention and group normalization for style consistency and a time-varying blend toward a uniform latent texture for dynamic alignment.
Key results
- FID of 74.58 vs 100.47 for TEXTure and 101.38 for Text2Tex; KID 2.89e-3 (best)
- User study: 50.00% preferred for visual quality and 46.28% for prompt alignment, the highest among methods
- Texture space sampling takes about 3 minutes per object on a single NVIDIA A10 GPU
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GenesisTex instead of TEXTure | Text-based 3D texture generation · GenesisTex evaluation set | FID (↓): 100.47 -> 74.58 | Helped |
| Use GenesisTex instead of TEXTure | Text-based 3D texture generation · GenesisTex evaluation set | KID (↓) (×10−3): 9.22 -> 2.89 | Helped |
| Use GenesisTex instead of TEXTure | Text-based 3D texture generation · User study | Visual Quality (%): 23.42 -> 50.00 | Helped |
| Use GenesisTex instead of TEXTure | Text-based 3D texture generation · User study | Align with Prompt (%): 24.86 -> 46.28 | Helped |
| Add Img2Img refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | FID (↓): 86.25 -> 82.05 | Helped |
| Add Img2Img refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | KID(↓) (×10−3): 3.79 -> 3.67 | Helped |
| Add inpainting refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | FID (↓): 86.25 -> 76.30 | Helped |
| Add inpainting refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | KID(↓) (×10−3): 3.79 -> 3.29 | Helped |
| Add both inpainting and Img2Img refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | FID (↓): 86.25 -> 74.58 | Helped |
| Add both inpainting and Img2Img refinement (ablation) | GenesisTex texture space sampling · Texture refinement evaluation | KID(↓) (×10−3): 3.79 -> 2.89 | Helped |
| Add style consistency (ablation) | GenesisTex texture space sampling · Multiple viewpoints | Multi-view style consistency | Helped |
| Add dynamic alignment (ablation) | GenesisTex texture space sampling with style consistency · Multiple viewpoints | Low-level multi-view consistency | Helped |
Limitations
- High memory cost of cross-view attention limits number of viewpoints, requiring inpainting and img2img post-processing
- TexFusion could not be compared directly because no public code was available; only figures extracted from its paper were shown
Experimental setup
Datasets: Objaverse ShapeNet Text2Mesh meshes Turbosquid Stanford 3D Scans Three D Scans
Benchmarks: 35-mesh / 80 (mesh, prompt) pair text-to-texture test set
Metrics: FID KID user study preference (visual quality, prompt alignment)
Baselines: Text2Mesh Latent-Paint Text2Tex TEXTure TexFusion
Builds on: DDIM ControlNet TexFusion Text2Tex TEXTure MVDream
Closest papers by meaning
- HIFA: High-fidelity Text-to-3D Generation with Advanced Diffusion Guidance (ICLR 2024 · similarity 0.90)
- SceneTex: High-Quality Texture Synthesis for Indoor Scenes via Diffusion Priors (CVPR 2024 · similarity 0.89)
- Single Mesh Diffusion Models with Field Latents for Texture Generation (CVPR 2024 · similarity 0.88)
- TexTailor: Customized Text-aligned Texturing via Effective Resampling (ICLR 2025 · similarity 0.88)
- ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models (CVPR 2024 · similarity 0.88)
- Text-Guided 3D Face Synthesis - From Generation to Editing (CVPR 2024 · similarity 0.87)
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.