Research map / 3D vision / 3D generation
Bootstrap3D: Improving Multi-view Diffusion Model with Synthetic Data
Zeyi Sun, Tong Wu, Pan Zhang, Yuhang Zang, Xiaoyi Dong, Yuanjun Xiong, Dahua Lin, Jiaqi Wang
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 |
|---|---|---|---|
| add 100k synthetic multi-view images without TTR (ablation) | Cap3D-captioned multi-view diffusion · Generated 3D objects | FID (PG2.5): 120.4 -> 134.6 | Hurt |
| add 100k synthetic images with TTR (ablation) | Cap3D-captioned multi-view diffusion · Multi-view images | FID (PG2.5): 101.3 -> 60.8 | Improved |
| add 100k synthetic images with TTR (ablation) | Cap3D-captioned multi-view diffusion · Multi-view images | CLIP-R Score: 77.9 -> 83.3 | Improved |
| add training timestep reschedule to synthetic data (ablation) | Cap3D + 100k synthetic · Generated 3D objects | CLIP-R Score: 71.2 -> 80.2 | Improved |
| add training timestep reschedule to synthetic data (ablation) | Cap3D + 100k synthetic · Multi-view images | FID (PG2.5): 92.0 -> 60.8 | Improved |
| replace Cap3D captions with MV-LLaVA dense recaptions (ablation) | Bootstrap3D + 100k synthetic · Multi-view images | CLIP-R Score: 83.3 -> 87.4 | Improved |
| replace Cap3D captions with MV-LLaVA dense recaptions (ablation) | Bootstrap3D + 100k synthetic · Multi-view images | FID (PG2.5): 60.8 -> 50.2 | Improved |
| increase synthetic data from 100k to 500k (ablation) | Bootstrap3D · Multi-view images | FID (PG2.5): 50.2 -> 42.4 | Improved |
| increase synthetic data from 100k to 500k (ablation) | Bootstrap3D · Multi-view images | CLIP-R Score: 87.4 -> 88.8 | Improved |
| increase synthetic data from 100k to 500k (ablation) | Bootstrap3D · Generated 3D objects | CLIP-R Score: 85.1 -> 86.3 | Improved |
| raise TTR threshold T for synthetic data (ablation) | Bootstrap3D · Multi-view images | view consistency vs prompt following: T=200 chosen | Mixed |
| use Bootstrap3D instead of Instant3D (unofficial) | text-to-multi-view · Real user prompts | CLIP-R Score: 77.0 -> 83.5 | Improved |
Closest papers by meaning
- IT3D: Improved Text-to-3D Generation with Explicit View Synthesis (AAAI 2024 · similarity 0.90)
- EfficientDreamer: High-Fidelity and Robust 3D Creation via Orthogonal-view Diffusion Priors (CVPR 2024 · similarity 0.90)
- Diffusion$^2$: Dynamic 3D Content Generation via Score Composition of Video and Multi-view Diffusion Models (ICLR 2025 · similarity 0.90)
- MVDream: Multi-view Diffusion for 3D Generation (ICLR 2024 · similarity 0.90)
- Vivid-ZOO: Multi-View Video Generation with Diffusion Model (NeurIPS 2024 · similarity 0.90)
- DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation (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.
