Research map / Image generation / Text-to-image diffusion
FlexiTex: Enhancing Texture Generation via Visual Guidance
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.
FlexiTex is a training-free texture generation framework that uses visual (image) guidance via IP-Adapter and direction-aware view prompts on a ControlNet-conditioned Stable Diffusion XL to produce consistent, detailed textures from text or image prompts.
Problem
Text-to-texture and image-to-texture generation for 3D meshes suffer from blurry or inconsistent textures, over-smoothing, and the Janus (multi-face) problem because abstract text prompts are ambiguous.
Approach
Convert the text prompt into an image with a text-to-image model and inject it as visual guidance through decoupled cross-attention (IP-Adapter) during synchronized multi-view denoising, and add view-specific direction prompts ("from <view> view") to reduce multi-face artifacts.
Key results
- Text-to-texture FID 81.72 and KID 8.78, the best among compared methods (SyncMVD FID 87.65), in 2.2 min per mesh.
- Image-to-texture CLIP Score 89.99 versus 87.43 for Paint3D, with FID 82.52 versus 103.44.
- Direction-aware adaptation lowers multi-face rate on animals from 48.00% to 21.50%, on objects from 8.00% to 3.00%, and on humans from 24.00% to 16.50%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FlexiTex instead of SyncMVD | Text-to-texture generation · Text-to-texture | FID: 87.65 -> 81.72 | Helped |
| Use FlexiTex instead of SyncMVD | Text-to-texture generation · Text-to-texture | KID: 9.53 -> 8.78 | Helped |
| Use FlexiTex instead of TexPainter | Text-to-texture generation · Text-to-texture | CLIP Score: 34.48 -> 34.31 | Hurt |
| Use FlexiTex instead of SyncMVD | Text-to-texture generation · Text-to-texture | Time: 1.8 min -> 2.2 min | Hurt |
| Use FlexiTex instead of Paint3D | Image-to-texture generation · Image-to-texture | FID: 103.44 -> 82.52 | Helped |
| Use FlexiTex instead of Paint3D | Image-to-texture generation · Image-to-texture | KID: 18.71 -> 11.97 | Helped |
| Use FlexiTex instead of Paint3D | Image-to-texture generation · Image-to-texture | CLIP Score: 87.43 -> 89.99 | Helped |
| Use FlexiTex instead of Paint3D | Image-to-texture generation · Image-to-texture | Time: 3.2 min -> 2.4 min | Helped |
| Add Direction-Aware Adaptation (ablation) | FlexiTex · Animal | Multi-face percentage: 48.00 % -> 21.50 % | Helped |
| Add Direction-Aware Adaptation (ablation) | FlexiTex · Object | Multi-face percentage: 8.00 % -> 3.00 % | Helped |
| Add Direction-Aware Adaptation (ablation) | FlexiTex · Human | Multi-face percentage: 24.00 % -> 16.50 % | Helped |
| Use Visual Guidance Enhancement instead of simple prompts (ablation) | FlexiTex · Texture generation | Detail richness | Helped |
Surprising findings
- On text-to-texture, FlexiTex's CLIP Score (34.31) is slightly below TexPainter (34.48), which the authors attribute to deviations introduced by the text-to-image module.
Limitations
- Generated results do not decouple lighting, causing artifacts in highlights or shadows.
- Minor dirty spots appear in areas where multi-view inputs are inconsistent or unobserved.
- Evaluation uses only 60 text-prompt and 60 image-prompt meshes, with FID/KID computed against samples from the diffusion model rather than real data.
Experimental setup
Datasets: Objaverse Objaverse-XL ShapeNet
Benchmarks: FID KID CLIP Score
Metrics: FID KID CLIP Score inference time multi-face percentage
Baselines: Latent-Paint Paint3D TexPainter Text2Tex SyncMVD TEXTure PGC-3D
Builds on: Stable Diffusion ControlNet IP-Adapter SyncMVD Score Distillation Sampling
Closest papers by meaning
- TexTailor: Customized Text-aligned Texturing via Effective Resampling (ICLR 2025 · similarity 0.87)
- Learning Visual Generative Priors without Text (CVPR 2025 · similarity 0.86)
- Generating compositional scenes via Text-to-image RGBA Instance Generation (NeurIPS 2024 · similarity 0.85)
- GenesisTex: Adapting Image Denoising Diffusion to Texture Space (CVPR 2024 · similarity 0.85)
- Flex3D: Feed-Forward 3D Generation with Flexible Reconstruction Model and Input View Curation (ICML 2025 · similarity 0.85)
- Tactile DreamFusion: Exploiting Tactile Sensing for 3D Generation (NeurIPS 2024 · similarity 0.84)
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.