README.md

metadata

title: Déjà View (DVLT)
emoji: 🔁
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
python_version: '3.12'
pinned: false
license: other
license_name: nvidia-internal-scientific-research-and-development-model-license
license_link: >-
  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-internal-scientific-research-and-development-model-license/
short_description: Multi-view 3D reconstruction with looping transformers
models:
  - nvidia/dvlt
tags:
  - 3d-reconstruction
  - depth-estimation
  - camera-pose
  - point-cloud
  - image-to-3d

🔁 Déjà View (DVLT)

Looping Transformers for Multi-View 3D Reconstruction — an interactive, unofficial demo.

Upload a video or a set of images and DVLT predicts per-view depth, camera poses and a 3D point cloud. It is a recurrent transformer: a shared attention block is looped K times, refining the prediction at each step — so the demo streams an intermediate point cloud after every step and you can watch the reconstruction converge.

⚠️ Unofficial demo. Not affiliated with or endorsed by NVIDIA. The nvidia/dvlt weights are released under the NVIDIA non-commercial research-and-evaluation license.

Links

• 📄 Paper: https://arxiv.org/abs/2605.30215
• 🌐 Project page: https://research.nvidia.com/labs/dvl/projects/dvlt/
• 🐙 Code: https://github.com/nv-tlabs/dvlt
• 🤗 Model: https://huggingface.co/nvidia/dvlt

Run locally

app.py is a single self-contained file with a PEP 723 header, so uv installs its deps and runs it:

uv run app.py    # http://localhost:7860

The @spaces.GPU decorator is a no-op off-Spaces, so it runs on CPU or a local GPU unchanged.

ZeroGPU note: ZeroGPU runs on Blackwell GPUs that need torch>=2.8, but the official dvlt package hard-pins torch==2.5.1. This repo vendors dvlt under packages/dvlt: repo/ is nv-tlabs/dvlt @ 265fe61 with a 3-line pin relaxation (torch/torchvision/pydantic), and wheels/ holds the prebuilt wheel (build.sh). It's installed from that wheel via the Space's own HTTPS resolve URL — not git or a local path, both of which fail on HF's Gradio build. See packages/dvlt/README.md.

Credits

DVLT by Burzio, Fischer, Elflein et al. (NVIDIA, UniMoRE, U. Toronto, ETH Zürich). Demo structure inspired by the VGGT-Ω and Depth-Anything-3 Spaces; example scenes are borrowed from those demos.