Research map / 3D vision / 3D generation
Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image
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.
Unique3D generates four orthographic multi-view RGB images and normal maps from a single image via fine-tuned Stable Diffusion models with multi-level upscaling, then reconstructs textured meshes in under 30 seconds with a mesh-based algorithm (ISOMER).
Problem
Single-image-to-3D mesh generation that is simultaneously high-fidelity, multi-view consistent, efficient, and generalizable, since SDS methods are slow and inconsistent while multi-view diffusion and LRM methods lack resolution and fine geometry.
Approach
Combine a multi-view RGB diffusion model and a normal diffusion model with progressive super-resolution (256→512→2048), then reconstruct the mesh directly with a mesh-based optimizer (ISOMER) that uses normal-map-based initialization, expansion regularization, and an ExplicitTarget per-vertex color/normal target that weights visible views by squared cosine to resolve multi-view inconsistency.
Key results
- On GSO, Unique3D reports PSNR 20.06 and LPIPS 0.107, outperforming InstantMesh (18.83, 0.128) and CRM (18.44, 0.137) on visual quality metrics.
- Unique3D achieves Chamfer Distance 0.0143 and F-Score 0.6696, with Volume IoU 0.5416 (best among methods on IoU and F-Score; CRM has the best CD at 0.0141).
- In a 30-volunteer user study, Unique3D scores 9.02 overall quality versus 8.43 for InstantMesh and 8.25 for CRM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ExplicitTarget to mesh reconstruction (ablation) | Unique3D ISOMER · GSO | F-Score: 0.6594 -> 0.6696 | Helped |
| add ExplicitTarget to mesh reconstruction (ablation) | Unique3D ISOMER · GSO | PSNR: 20.0383 -> 20.0611 | Helped |
| add ExplicitTarget to mesh reconstruction (ablation) | Unique3D ISOMER · GSO | LPIPS: 0.1129 -> 0.1070 | Helped |
| add ExplicitTarget to multi-view colorization (ablation) | Unique3D · multi-view mesh colorization | texture artifacts | Helped |
| add expansion regularization (ablation) | Unique3D ISOMER · mesh reconstruction | collapse incidence | Helped |
| use multi-level super-resolution (ablation) | Unique3D · high-resolution mesh generation | detail resolution | Helped |
| use fast mesh initialization instead of a sphere (ablation) | Unique3D ISOMER · mesh reconstruction | reconstruction quality | Mixed |
| replace Wonder3D reconstruction with ISOMER | Wonder3D · GSO | Chamfer Distance: 0.0261 -> 0.0244 | Hurt |
Surprising findings
- Replacing Wonder3D's reconstruction with ISOMER improves its PSNR from 18.09 to 18.61 while being faster, showing the reconstruction module is usable as a plug-in.
Limitations
- Multi-view prediction may give less satisfactory results for skewed or non-perspective inputs.
- The geometric coloring algorithm does not currently support texture maps.
- Out-of-distribution in-the-wild inputs cause multi-view inconsistency that the method only partly addresses.
Experimental setup
Datasets: Objaverse GSO
Benchmarks: GSO
Metrics: PSNR SSIM LPIPS CLIP-Similarity Chamfer Distance Volume IoU F-Score
Baselines: One-2-3-45 OpenLRM Wonder3D InstantMesh CRM SyncDreamer GRM
Builds on: Stable Diffusion Image Variations ControlNet-Tile Zero123 Real-ESRGAN LGM
Closest papers by meaning
- One-2-3-45++: Fast Single Image to 3D Objects with Consistent Multi-View Generation and 3D Diffusion (CVPR 2024 · similarity 0.92)
- Cycle3D: High-quality and Consistent Image-to-3D Generation via Generation-Reconstruction Cycle (AAAI 2025 · similarity 0.91)
- Wonder3D: Single Image to 3D using Cross-Domain Diffusion (CVPR 2024 · similarity 0.90)
- Ouroboros3D: Image-to-3D Generation via 3D-aware Recursive Diffusion (CVPR 2025 · similarity 0.89)
- IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation (ICML 2024 · similarity 0.89)
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · similarity 0.89)
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.