AppliedIntuitionResearch/BuildRome

Building Rome from a Single Image

Building Rome generates a complete metric 3D scene mesh — visible surfaces and plausible occluded geometry — from a single photograph, indoors and outdoors. It fine-tunes the TRELLIS.2 object generator to work on scenes: a monocular depth estimate (MoGe-3) lifts image features onto the observed surfaces, the scene is covered with chunks that grow with distance from the camera, and the chunks are generated one at a time with latent outpainting and stitched into one mesh.

⚠️ Research release. This model is released for research purposes under the terms below. It is not a production system and does not include any Applied Intuition proprietary data, product code, or production checkpoints.

Built with DINOv3.

Model Details

Developed by	Applied Intuition — AI Research
Model type	LoRA adapters and depth-conditioning branches for the TRELLIS.2 sparse-structure (SS) and shape (SLat) flow transformers
Paper	Building Rome from a Single Image
Code	https://github.com/Applied-Intuition-Open-Source/BuildRome
Project page	https://build-rome.github.io/
Base model	microsoft/TRELLIS.2-4B (MIT), downloaded separately
Training data	Indoor scenes from SAGE-10k and Infinigen, and about 4,000 synthetic outdoor scenes (see the paper)
License (weights)	CC BY-NC 4.0
Contact	GitHub Issues on the code repo

Checkpoints

The files hold only the fine-tuned parameters; the inference code adds them to the public TRELLIS.2 base weights.

Checkpoint	Description	Size	Link
ss_delta.safetensors	Sparse-structure (occupancy) model: LoRA r32 adapters and depth-lift / clearance branches, 115M params.	438 MiB	ss_delta.safetensors
slat_delta.safetensors	Shape model: LoRA r32 adapters and depth-lift branch, 66M params.	251 MiB	slat_delta.safetensors

SHA-256 checksums are in SHA256SUMS.

Intended Use & Limitations

Intended use: research on single-image 3D scene reconstruction and generation.

Out of scope / limitations:

• Outputs are geometry only (no texture). Metric scale and camera come from the MoGe-3 depth estimate, so depth errors carry into the scene.
• Occluded regions are plausible completions, not reconstructions of the true scene.
• Meshes are not guaranteed to be watertight and can contain holes, as with TRELLIS.2.
• Requires a CUDA GPU; a 98 m street scene (4 depth bands, 15 chunks) peaks at about 16 GiB and takes about 6 min on one A100.

Commercial use is not permitted under CC BY-NC 4.0.

How to Use

from build_rome import ScenePipeline

pipe = ScenePipeline.from_pretrained("AppliedIntuitionResearch/BuildRome")
scene = pipe("photo.jpg")
scene.save("results/photo")  # scene.glb, layout_bands.json, layout_topview.png, spiral.mp4

Command line:

hf download AppliedIntuitionResearch/BuildRome --local-dir checkpoints
python inference/image_to_scene.py --image photo.jpg \
    --ss-ckpt checkpoints/ss_delta.safetensors --slat-ckpt checkpoints/slat_delta.safetensors --out results/

The pipeline also downloads TRELLIS.2-4B, TRELLIS-image-large, DINOv3 ViT-L/16 (gated: accept its license on Hugging Face first) and MoGe-3. Full installation and usage instructions: see the GitHub repository.

License

The weights in this repository are released under CC BY-NC 4.0, non-commercial use only (full text in LICENSE). They are used together with separately downloaded models that keep their own terms:

• TRELLIS.2-4B and TRELLIS-image-large: MIT.
• DINOv3: the DINOv3 License. Built with DINOv3.
• MoGe-3: MIT.

Training data: SAGE-10k (Apache-2.0), Infinigen scenes (BSD-3-Clause generator), and synthetic outdoor scenes produced with the pipeline described in the paper. No training data is included in this repository.

Citation

@article{yenphraphai2026rome,
  title   = {Building Rome from a Single Image},
  author  = {Yenphraphai, Jiraphon and Li, Fang and Xu, Tianshuo and Meng, Depu and Herau, Quentin and Hu, Yihan and Yeh, Raymond A. and Zhan, Wei},
  journal = {arXiv preprint arXiv:2610.08790},
  year    = {2026}
}

Acknowledgments

This model builds on TRELLIS.2 and TRELLIS (Microsoft, MIT), MoGe (Microsoft, MIT) and DINOv3 (Meta, DINOv3 License). Training uses scenes from SAGE-10k (NVIDIA) and Infinigen (Princeton). We thank the authors for making their work available.