MV-SAM3D is a multi-view 3D reconstruction framework that extends SAM 3D Objects to leverage observations from multiple viewpoints. It supports both single-object and multi-object generation, and is designed to produce more stable geometry, texture, and scene-level consistency.

• arXiv: https://arxiv.org/abs/2603.11633

Please follow the environment setup from:

• SAM 3D Objects
• Depth Anything 3

scene/
├── images/
│   ├── 0.png
│   ├── 1.png
│   └── ...
├── object_a/
│   ├── 0.png
│   ├── 1.png
│   └── ...
├── object_b/
│   └── ...
└── ...

Mask files are RGBA PNG where alpha indicates foreground.

Single-View (View 3)	Single-View (View 6)	MV-SAM3D
Input Image	Input Image	Input Images
↓ 3D Reconstruction ↓
Single-view baseline.	Single-view baseline.	Better multi-view consistency.

SAM 3D (single-view)	MV-SAM3D w/o Pose Optimization	MV-SAM3D (full)
Shape and pose are often unstable.	Multi-view improves object quality.	Improved overall scene alignment.

python run_inference_weighted.py \
  --input_path ./data/example \
  --mask_prompt stuffed_toy \
  --da3_output ./da3_outputs/example/da3_output.npz

python run_inference_weighted.py \
  --input_path ./data/desk_objects0 \
  --mask_prompt keyboard,speaker,mug,stuffed_toy \
  --da3_output ./da3_outputs/desk_objects0/da3_output.npz \
  --merge_da3_glb \
  --run_pose_optimization

For single-object inference (run_inference_weighted.py), key defaults are:

• Stage 1 weighting: enabled (stage1_entropy_alpha=30.0)
• Stage 2 weighting: enabled (stage2_weight_source=entropy)
• Stage 2 alpha defaults: stage2_entropy_alpha=30.0, stage2_visibility_alpha=30.0

python preprocessing/build_mvsam3d_dataset.py \
  --input data/your_scene \
  --objects keyboard,speaker,mug,stuffed_toy

python scripts/run_da3.py \
  --image_dir ./data/your_scene/images \
  --output_dir ./da3_outputs/your_scene

@article{li2026mv,
  title={MV-SAM3D: Adaptive Multi-View Fusion for Layout-Aware 3D Generation},
  author={Li, Baicheng and Wu, Dong and Li, Jun and Zhou, Shunkai and Zeng, Zecui and Li, Lusong and Zha, Hongbin},
  journal={arXiv preprint arXiv:2603.11633},
  year={2026}
}

We thank the authors of SAM 3D Objects and Depth Anything 3 for their excellent work!!!

Please refer to LICENSE for usage terms.