---
title: Déjà View
emoji: 👁
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.31.0
python_version: '3.12'
app_file: app.py
license: apache-2.0
---
# Déjà View

DVLT reconstructs camera poses and 3D geometry from images or video.
The hosted app serves DVLT only, with 8–16 inference steps and cached
16-step examples. Visualization controls are under Advanced.
Model selection is available only from the command line, not in the GUI.
The **Filter sky** control is enabled by default and available under Advanced.

Reconstruction displays and logs **Network forward pass** time with CUDA
synchronized before and after the network call. Loading, preprocessing,
prediction postprocessing, sky filtering and file export are excluded.
Displayed throughput FPS is input frame count divided by network forward seconds.
Visualization-only changes retain the last timing; cached examples do not
run the network or display a newly measured time.

The intended deployment uses **Protected** visibility: the app is publicly
accessible at its Hugging Face app URL, while this source repository stays
private. Visibility is a Space setting, not controlled by this README.

## Uploads and temporary results

Live inference accepts JPEG, PNG or WebP images and supported video files,
up to 300 images or sampled video frames, at most
16 megapixels per frame. Uploads are limited to 100 MB per file at the
HTTP upload boundary and 256 MiB total per submitted collection. Videos
are sampled at 0.5–30 FPS, with no limit on the original video frame count.
Video decoding has a 30-second timeout.

Uploads, previews and reconstruction downloads use temporary server storage.
Gradio cache cleanup is configured to remove files older than one hour,
checked every five minutes while the server is running. In-memory session
state has a one-hour expiry. These application settings do not establish
Hugging Face platform-log, backup or infrastructure retention. Those must
be confirmed separately before any broader retention promise.

Results include downloadable GLB and RRD files. Fixed examples also expose
input previews, depth previews and the camera/depth files listed in their
manifest. Cached examples are persistent demonstration assets, separate
from visitor uploads. Do not upload confidential or sensitive content.

## Licensing and model identity

The code is primarily Apache 2.0 with third-party exceptions documented in
`THIRD_PARTY_LICENSES.md` and `LICENSES/`. The model weights have separate
NVIDIA non-commercial research/evaluation terms; the code license does not
license the model or establish the approved terms of this hosted service.
The hosted model is pinned to `nvidia/dvlt` revision
`20f919875a00ee3ab2e840f2f31580f4652b7c5e`.

Service terms and privacy notices must reflect the approved release review.

[Project](https://research.nvidia.com/labs/dvl/projects/dvlt/) ·
[Code](https://github.com/nv-tlabs/dvlt)