Research map / 3D vision / 3D generation
BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing
Jiacheng Chen, Ramin Mehran, Xuhui Jia, Saining Xie, Sanghyun Woo
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 dual-stream architecture (ablation) | 3DIT one-stream baseline · Objectron (ablation) | FID: 6.14 -> 3.52 | Improved |
| add dual-stream architecture (ablation) | 3DIT one-stream baseline · Objectron (ablation) | Object PSNR: 13.88 -> 15.90 | Improved |
| add Depth Pro and SAM2 depth and segmentation to the source stream (ablation) | Dual-stream · Objectron (ablation) | FID: 3.52 -> 3.74 | No clear effect |
| add Blender renders as 3D-grounded control signals in both streams (ablation) | Dual-stream · Objectron (ablation) | FID: 3.52 -> 2.93 | Improved |
| add source masking training (ablation) | DS + Blender · Objectron (ablation) | FID: 2.93 -> 2.64 | Improved |
| add simulated object jittering training (ablation) | DS + Blender + source masking · Objectron (ablation) | FID: 2.64 -> 3.25 | Hurt |
| initialize WOD training with MOVi-E pretrained checkpoint (ablation) | BlenderFusion · WOD | object rotation | Improved |
| use BlenderFusion instead of Object 3DIT | Stable Diffusion v2.1 compositor · MOVi-E (standard video setup) | Object-level PSNR: 14.06 -> 18.90 | Improved |
| use BlenderFusion instead of Neural Assets | Stable Diffusion v2.1 compositor · WOD (standard video setup) | Frame-level FID: 15.39 -> 10.02 | Improved |
| use BlenderFusion instead of Neural Assets in human study | BlenderFusion · Human evaluation (overall) | win rate: 87.04 vs 6.40 | Improved |
| MOVi-E pre-training for object rotation failures on Objectron (ablation) | BlenderFusion · Objectron | object rotation | No clear effect |
| keep NA at 256x256 instead of higher resolution | Neural Assets · Objectron/WOD | quantitative metrics | 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.
- Neural Assets — Neural Assets: 3D-Aware Multi-Object Scene Synthesis with Image Diffusion Models (NeurIPS 2024) · better on 10 of 10 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
- Zero-Shot Depth Aware Image Editing with Diffusion Models (ICCV 2025 · similarity 0.89)
- Generative Image Layer Decomposition with Visual Effects (CVPR 2025 · similarity 0.88)
- ProEdit: Simple Progression is All You Need for High-Quality 3D Scene Editing (NeurIPS 2024 · similarity 0.88)
- BlenderGym: Benchmarking Foundational Model Systems for Graphics Editing (CVPR 2025 · similarity 0.88)
- SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 2025 · similarity 0.87)
- PS-Diffusion: Photorealistic Subject-Driven Image Editing with Disentangled Control and Attention (CVPR 2025 · similarity 0.87)
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.
