Research map / 3D vision / 3D generation
MVPaint: Synchronized Multi-View Diffusion for Painting Anything 3D
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.
MVPaint is a three-stage text-to-texture framework (synchronized multi-view diffusion, learning-free 3D point-cloud inpainting, and UV refinement with seam smoothing) that produces consistent, seamless 2K textures for 3D meshes.
Problem
Text-to-texture generation for 3D meshes suffers from local discontinuities, cross-view inconsistency, Janus problems, and heavy dependence on UV unwrapping quality.
Approach
Synchronize multi-view diffusion by fusing decoded multi-view images into a shared UV map (weighted by view-normal cosine), inpaint missing texture in 3D point-cloud space rather than UV space, and refine in UV space with spatial-aware seam smoothing so results are robust to UV unwrapping.
Key results
- On the Objaverse T2T benchmark, MVPaint achieves FID 20.89 vs. 25.28 for the best prior method (Paint3D) and KID 3.45 vs. 5.19.
- On the GSO T2T benchmark, MVPaint reaches FID 20.02 and KID 3.12, best on both, and the highest user-study ratings (overall 4.13, seamless 4.51, consistency 4.21).
- Ablation shows removing multi-view synchronization raises FID from 20.89 to 21.42 and removing the multi-view diffusion model raises it to 27.63.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVPaint rather than Paint3D | Text-to-texture generation · Objaverse T2T benchmark | FID: 25.28 -> 20.89 | Helped |
| Use MVPaint rather than Paint3D | Text-to-texture generation · Objaverse T2T benchmark | KID: 5.19 -> 3.45 | Helped |
| Use MVPaint rather than TEXTure | Text-to-texture generation · Objaverse T2T benchmark | CLIP: 20.30 -> 19.87 | Hurt |
| Use MVPaint rather than TEXTure | Text-to-texture generation · GSO T2T benchmark | FID: 24.76 -> 20.02 | Helped |
| Use MVPaint rather than SyncMVD | Text-to-texture generation · GSO T2T benchmark | KID: 5.37 -> 3.12 | Helped |
| Use MVPaint rather than TEXTure | Text-to-texture generation · GSO T2T benchmark | CLIP: 23.44 -> 23.25 | Hurt |
| Add multi-view synchronization (ablation) | MVPaint · Objaverse T2T benchmark | FID: 21.42 -> 20.89 | Helped |
| Add the T2MV diffusion model (ablation) | MVPaint · Objaverse T2T benchmark | FID: 27.63 -> 20.89 | Helped |
| Add geometry-guided refinement (ablation) | MVPaint · Objaverse T2T benchmark | FID: 21.17 -> 20.89 | Helped |
| Add 3D inpainting (ablation) | MVPaint · Objaverse T2T benchmark | FID: 20.91 -> 20.89 | Helped |
| Add seam smoothing (ablation) | MVPaint · Objaverse T2T benchmark | FID: 20.82 -> 20.89 | Hurt |
Surprising findings
- 3D inpainting and seam smoothing have minimal effect on FID/KID (within ±0.1 and below 0.2 KID), though they remove visible texture artifacts.
Limitations
- Relies on Janus-free multi-view prior and multi-stage pipeline; computational cost not reported.
- Quantitative ablations show small metric differences for inpainting and seam smoothing, so their benefit is mainly qualitative.
Experimental setup
Datasets: Objaverse Paint3D test set GSO
Benchmarks: Objaverse T2T benchmark GSO T2T benchmark
Metrics: FID KID CLIP score user study overall quality seamless consistency
Baselines: TEXTure Paint3D SyncMVD
Builds on: MVDream SyncMVD ControlNet SDXL Paint3D
Closest papers by meaning
- VideoScene: Distilling Video Diffusion Model to Generate 3D Scenes in One Step (CVPR 2025 · similarity 0.85)
- MVDream: Multi-view Diffusion for 3D Generation (ICLR 2024 · similarity 0.85)
- Enhancing Novel View Synthesis via Geometry Grounded Set Diffusion (NeurIPS 2026 · similarity 0.83)
- HarmonyView: Harmonizing Consistency and Diversity in One-Image-to-3D (CVPR 2024 · similarity 0.83)
- Era3D: High-Resolution Multiview Diffusion using Efficient Row-wise Attention (NeurIPS 2024 · similarity 0.82)
- RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing (CVPR 2025 · similarity 0.82)
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.