Research map / 3D vision / Face and shape
FlashTex: Fast Relightable Mesh Texturing with LightControlNet
Kangle Deng, Timothy Omernick, Alexander B Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, Maneesh Agrawala
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace distilled encoder with original SD encoder (ablation) | FlashTex · Objaverse test set | runtime (mins): 4 -> 8 | Hurt |
| add distilled Stable Diffusion image encoder (ablation) | FlashTex · Objaverse test set | FID: 60.34 -> 62.67 | No clear effect |
| add multi-view visual prompting (2x2 grid of canonical views) (ablation) | FlashTex stage 1 · Objaverse test set | FID: 74.23 -> 62.67 | Improved |
| add multi-view visual prompting (ablation) | FlashTex · Objaverse test set | runtime (mins): 19 -> 4 | Improved |
| use all three material bases in the conditioning image (ablation) | LightControlNet · Objaverse test set | FID: 72.13 -> 62.67 | Improved |
| use four canonical views instead of two (ablation) | FlashTex stage 1 · Objaverse test set | FID: 67.43 -> 62.67 | Improved |
| add top and bottom views (six views instead of four) (ablation) | FlashTex stage 1 · Objaverse test set | FID: 62.67 -> 70.14 | Hurt |
| train LightControlNet on more directional lighting versus less directional (ablation) | LightControlNet · Objaverse test set | FID: 70.14 -> 62.67 | Improved |
| use LightControlNet instead of depth ControlNet | FlashTex · Objaverse test set | KID (x10^-3): 3.96 -> 2.69 | Improved |
| use depth ControlNet instead of LightControlNet when relighting is not needed | FlashTex · Objaverse test set | runtime (mins): 4 -> 2 | Improved |
| use Variational Score Distillation instead of SDS (ablation) | FlashTex stage 2 · Objaverse | texture quality | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Text2tex — Text2Tex: Text-driven Texture Synthesis via Diffusion Models (ICCV 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering (CVPR 2024 · similarity 0.86)
- TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features (CVPR 2025 · similarity 0.86)
- GOATex: Geometry & Occlusion-Aware Texturing (NeurIPS 2025 · similarity 0.85)
- Paint3D: Paint Anything 3D with Lighting-Less Texture Diffusion Models (CVPR 2024 · similarity 0.85)
- SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement (CVPR 2025 · similarity 0.85)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · 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.
