AI Depth Map
Select a video clip and open Properties > Tracking > AI Depth Map.
- Download depth model loads Depth Anything V2 Small directly from Hugging Face (99,060,839 bytes, FP32). The pinned model is SHA-256 verified and stored in a dedicated browser Cache Storage cache. A corrupt cached file is discarded. Cache failures leave the model usable for the session. No video is uploaded.
- Estimate current frame shows relative inverse depth: white is near, black is far. Start live depth follows the playhead until stopped or the panel closes. The source timestamp and inference time are displayed; this is an asynchronous analysis preview, not guaranteed full-framerate playback.
- Fast uses a 280-pixel long edge; Detailed uses 518. Input dimensions are multiples of 14. The source aspect ratio is retained when displaying/exporting.
- Range smoothing stabilizes the 2nd/98th percentile display range; it does not smear moving geometry. Seeking backwards, large jumps and range discontinuities reset history. Polarity can be inverted.
- Bake depth video to Media deterministically analyzes the chosen source range at 10, 15 or 30 fps, then imports a silent grayscale H.264 MP4. The range is bounded to 120 seconds. The map starts at output time zero corresponding to Source start. It excludes clip transforms, effects, speed changes and audio; apply matching timing/transforms when aligning it with an edited clip. Once baked, it uses ordinary media playback, project storage and export, without ML.
- Stop depth cancels decoding/download/inference and keeps previously imported results. Unload model releases the worker/GPU while retaining the download; Clear model cache also removes its downloaded weights.
Runtime and provenance
Section titled “Runtime and provenance”src/services/depthEstimation/ separates verified model loading, worker ownership,
relative-depth math and sequential video baking. Runtime objects remain outside
project data and stores. The HMR-preserved runtime owns one worker and permits one
inference at a time; no playback-sized job queue is accumulated. Worker requests
have a two-minute timeout. Model execution uses the existing Transformers.js
runtime, with WebGPU preferred and an explicitly labelled WASM fallback.
Transformers.js runtime assets load from its versioned jsDelivr distribution;
model cache availability alone therefore does not guarantee a fully offline cold
start. Backend performance is device dependent.
Model: onnx-community/depth-anything-v2-small,
revision 4472b7362082ad9968fee890ca0f1e5aca36b93d, file onnx/model.onnx,
SHA-256 afb6a5c28f3b6bf1618c6e43f02073ef9dfdc70e937502d51603e57b0a1df10c.
The source model and conversion identify Apache-2.0; see
third-party notices.
Boundaries
Section titled “Boundaries”This produces relative monocular depth, not distances in meters or a complete 3D mesh. The analysis preview is in source space. Face Cable collision/shadows and Voxel Relief depth sampling are separate consumers and are not yet connected. The depth-aware Voxel Relief orchestration plan remains open; this feature does not mark its broader renderer/export/3D scope complete.