---
title: ZDPShift · Unmuzzling Stereo
emoji: 🎬
colorFrom: red
colorTo: blue
sdk: gradio
app_file: app.py
pinned: false
license: mit
---

# Unmuzzling Stereo — signed disparity demo

Six stereo backbones (RAFT-Stereo, IGEV-Stereo, FoundationStereo,
DynamicStereo, BiDAStereo, StereoAnyVideo), each runnable in corrected form
(signed disparity, content on both sides of the screen plane) and optionally
side by side with the released zero-shot model, which cannot represent
negative disparity. Image models are fine-tuned on translated SceneFlow only;
video models on translated SceneFlow + Dynamic Replica, matching the data of
their released baselines. The ZDPShift benchmark itself is never trained on.

Code: [github.com/shijianjian/ZDPShift](https://github.com/shijianjian/ZDPShift) ·
Weights: [shijianjian/ZDPShift](https://huggingface.co/shijianjian/ZDPShift) ·
Project page: [shijianjian.github.io/ZDPShift](https://shijianjian.github.io/ZDPShift/).
Rendered examples come from Blender Studio open movies (CC-BY) at the largest
ZDP shift, where most of the scene sits behind the screen plane. The film
examples (Jurassic World, Avatar, Galapagos, Kung Fu Panda) are delivered
stereo frames from public trailers, shown for research illustration.

Local run: `ZDP_WEIGHTS_ROOT=/path/to/weights python app.py`