Research map / 3D vision / 3D generation
VFusion3D: Learning Scalable 3D Generative Models from Video Diffusion Models
Junlin Han, Filippos Kokkinos, Philip Torr
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 |
|---|---|---|---|
| use VFusion3D trained on synthetic multi-view data instead of OpenLRM or LGM | VFusion3D · GSO (996 objects) | PSNR: 14.21 -> 20.89 | Improved |
| increase 3D fine-tuning data for Emu Video from 10K to 100K (ablation) | Emu Video fine-tuning · good rate of generated multi-view videos | good rate: 44.5% -> 61.3% | Improved |
| add multi-stage training recipe (ablation) | VFusion3D · 500K synthetic data | LPIPS: 0.206 -> 0.168 | Improved |
| remove pixel-level (MSE) image supervision (ablation) | VFusion3D · 500K synthetic data | CLIP text similarity: 0.249 -> 0.257 | Improved |
| add opacity supervision (ablation) | VFusion3D · 500K synthetic data | CLIP image similarity: 0.798 -> 0.802 | Improved |
| add camera noise injection (ablation) | VFusion3D · 500K synthetic data | CLIP text similarity: 0.256 -> 0.252 | No clear effect |
| fine-tune VFusion3D with more 3D data (none to 100K) (ablation) | VFusion3D · test 3D models | LPIPS: 0.160 -> 0.143 | Improved |
| select 3D fine-tuning data by aesthetic score instead of random (ablation) | VFusion3D · test 3D models | LPIPS: 0.147 vs 0.146 (random) | No clear effect |
| use 32 rendered views instead of 16 for 3D fine-tuning (ablation) | VFusion3D · test 3D models | SSIM: 0.838 -> 0.839 | No clear effect |
| train on synthetic multi-view data instead of 150K real 3D data (ablation) | VFusion3D · text-to-3D generalization | CLIP image similarity: 0.631 -> 0.839 | Improved |
| combine synthetic pre-training with 3D data fine-tuning (ablation) | VFusion3D · test 3D models | LPIPS: 0.160 -> 0.143 | Improved |
| increase synthetic dataset size from 100K to 2.7M (ablation) | VFusion3D · scaling study | LPIPS and CLIP image similarity | Improved |
Later papers that compared with this method
2 later accepted papers report VFusion3D as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Flex3D: Feed-Forward 3D Generation with Flexible Reconstruction Model and Input View Curation (ICML 2025 · better on 4 of 4 shared results)
- MVBoost: Boost 3D Reconstruction with Multi-View Refinement (CVPR 2025 · better on 3 of 3 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
- WildFusion: Learning 3D-Aware Latent Diffusion Models in View Space (ICLR 2024 · similarity 0.90)
- Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models (ICLR 2025 · similarity 0.89)
- MVDream: Multi-view Diffusion for 3D Generation (ICLR 2024 · similarity 0.89)
- SV4D: Dynamic 3D Content Generation with Multi-Frame and Multi-View Consistency (ICLR 2025 · similarity 0.89)
- IT3D: Improved Text-to-3D Generation with Explicit View Synthesis (AAAI 2024 · similarity 0.89)
- GenFusion: Closing the Loop between Reconstruction and Generation via Videos (CVPR 2025 · 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.
