Try it · Modes · Motion-control guide · 中文
Depth Video Studio turns any workout or performance video into the control inputs AI video generators need — grayscale depth maps, pose skeletons, 478-point face clouds — entirely inside your browser. No install, no server, no upload, original audio preserved.
Built for motion control in Seedance 2.0, Kling, Runway, Viggle and for fitness content creators who are tired of watermark apps, GPU setup guides, and tools that quietly upload their footage.
A 12-second vertical fitness clip in, motion-control MP4s out — one analysis pass, then every mode re-exports in real time:
Left: live depth conversion · Right: depth frame vs. black-background skeleton frame (the control format)
| 12 s clip, 24 fps, 288 frames | Analysis | Export |
|---|---|---|
| Chrome/Edge + WebGPU (discrete GPU) | ~15–30 s | real-time |
| Chrome/Edge + WebGPU (integrated) | ~30–60 s | real-time |
| WASM fallback | 1–5 min | real-time |
Switching modes reuses the analysis cache — export depth, then the skeleton, each additional export costs only the clip's real-time duration. ZIP archives carry UTF-8 filenames (fine in Finder and Windows Explorer); if you extract from a terminal, prefer ditto -x -k file.zip out/ or 7-Zip — the ancient Info-ZIP unzip mangles non-ASCII names.
The "AI workout clone" workflow needs a depth or skeleton video next to your reference clip. Today people get it by:
- running Python repos with CUDA environments and 2 GB of downloads, or
- uploading their footage to random web tools, or
- paying for watermark-locked mobile apps at 480p.
This project is a single HTML file that does the whole job locally: Depth Anything V2 for depth, MediaPipe for pose (33 keypoints) and face (478 points incl. iris), WebGPU for speed with automatic WASM fallback, and MediaRecorder for MP4/WebM export with the original audio muxed in.
Online (recommended): open the hosted demo → https://beatapi.github.io/depth-video-studio/
Offline / single file: download index.html and double-click it.
From this repo:
git clone https://github.com/BeatAPI/depth-video-studio.git
cd depth-video-studio
python3 serve.py # or: start-server.bat (Windows) / start-server.command (macOS)Requirements: a modern browser — Chrome/Edge 113+ for WebGPU acceleration (Safari 18+/Firefox work via WASM). First run downloads ~65 MB of models, cached by the browser for offline use afterwards. If Hugging Face is unreachable, the app retries via hf-mirror.com automatically. No Python, no GPU drivers, no build step.
| Mode | Output | Use it for |
|---|---|---|
| Depth map | Grayscale mono-depth, near = bright (Inferno colormap optional) | depth control video (recommended default) |
| Pose skeleton | 33 keypoints over the video, or on pure black | pose control / OpenPose-style input |
| Depth + pose | depth background with skeleton overlay | QA and content B-roll |
| Face 478-point cloud | full face mesh incl. iris highlights | expression reference |
| Everything | depth + skeleton + face combined | visual comparison |
Workflow niceties: drag-and-drop, live per-frame preview during analysis, two-phase progress, cancel anytime, temporal EMA smoothing to kill per-frame depth jitter, analysis fps 6–30, bitrate 4–16 Mbps, a bilingual UI (中文 / English), and batch mode — queue up to 10 clips with one mode/parameter set, download each result or grab them all as a single ZIP:
Short version — full guide in docs/seedance-workflow.md:
- Use grayscale depth + your character image as the reference/first frame; keep aspect ratios matched.
- Temporal smoothing: Light. Bitrate: 8–16 Mbps — artifacts in the control video leak into the generation.
- For limb-precise transfer, export the skeleton on pure black instead.
Also pairs well with Kling/Vidu/Wan motion control, Runway-style references, and ComfyUI depth pipelines when you want a zero-install path.
All inference runs locally in your browser. There is no backend, no analytics, and no upload — the entire app is one readable index.html, audit it yourself.
- Batch queue — up to 10 clips, one click, ZIP download
- Multi-person pose (
numPoses > 1) - Hand landmarks (21 × 2)
- Larger depth model option (DA-V2 large / Depth Pro)
- PNG-sequence export for ComfyUI
- Folder batch via File System Access API
Contributions welcome — the whole app is a single readable file, and good first issues are tagged.
Code: MIT. Models run on demand from their official CDNs under their own licenses (Depth Anything V2 — Apache-2.0; MediaPipe — Apache-2.0); this repo redistributes no weights.
由 BeatAPI 维护 · beatapi.io — 音乐视频与产品广告的异步 AI 视频 API · 中文文档