EnVision-Research

EnVision-Research / DVD

Public

DVD: Deterministic Video Depth Estimation with Generative Priors

79
7
100% credibility
Found Mar 13, 2026 at 71 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

DVD is an academic tool for creating accurate, stable depth maps from monocular videos by adapting generative video models into deterministic predictors.

How It Works

1
📰 Discover DVD

You hear about DVD, a smart tool that turns everyday videos into detailed 3D depth maps, perfect for seeing the hidden structure in motion.

2
📥 Get the package

Download the simple starter kit to your computer so everything is ready to go.

3
🧠 Add the model

Fetch the trained brain of the tool from a trusted sharing site to make it think.

4
🎥 Test with demos

Drop in the included sample videos and press go to see depth maps appear instantly.

5
✨ Magic happens

Watch as your videos transform into smooth, flicker-free depth videos with sharp edges and long stability.

6
📱 Use your videos

Swap in your own clips, adjust size if needed, and generate pro-level depth for any adventure.

✅ Depth mastery

Celebrate having precise, stable 3D depth from videos to explore, analyze, or create with ease.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 71 to 79 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is DVD?

DVD delivers deterministic depth estimation for videos and images, converting raw footage into precise depth maps without the hallucinations plaguing generative models. It adapts pre-trained video diffusion models like WanV2.1 into fast, single-pass regressors, ensuring temporal stability even on long clips. Built in Python with Hugging Face integration, users run simple bash scripts—like `openworld.sh` for demos or `video.sh` for benchmarks—to generate depth outputs on custom data.

Why is it gaining traction?

Unlike stochastic diffusion baselines, DVD strips away noise for repeatable results, hitting SOTA on boundary precision (Recall/F1) while using 163x less training data (367K frames). Its Global Affine Coherence enables seamless long-video stitching with minimal drift, and inference skips slow ODE solvers. Developers dig the plug-and-play pre-trained weights and configs for KITTI, ScanNet, or Bonn datasets.

Who should use this?

Computer vision researchers benchmarking depth on video datasets like KITTI Eigen or NYU Depth V2. Robotics engineers needing stable, real-time scene depth for SLAM or navigation. AR/VR devs prototyping 3D reconstruction from handheld footage.

Verdict

Worth forking for non-commercial depth experiments—solid README, HF models, and eval scripts make it accessible despite 46 stars and 1.0% credibility score. Still early (fresh arXiv), so expect tweaks for production; pair with Depth Anything for hybrids.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.