Transformers depth lab

What you are looking at. The large frame is the depth map inferred from your camera or image (Transformers.js depth-estimation pipeline). Pale / bright usually means closer in relative terms—exact mapping depends on the model. It is not a live duplicate webcam view unless you enable RGB underlay. Start the camera, wait for "Ready", then watch the depth update (rate depends on CPU/GPU and settings below).

Source

Or load image

HTTPS or localhost. First model download can take a while from the Hub. Optimized for desktop Chrome. If you load a file image, it overrides the webcam for depth until you clear the file.

Model

Preset (click card for Hub page)

Custom model id (Hub)

Only models published with ONNX for Transformers.js work here. Filtered list: Hub · depth-estimation · transformers.js. Heavier models need WebGPU or fewer updates per second.

Runtime

Device dtype

WebGPU often needs fp16 or fp32. If init fails, switch back to WASM + q8.

Inference

Depth on CPU (WASM) is expensive. Use a smaller max side, higher throttle, or WebGPU if your GPU supports it.

Max input side (px) — 384 Min ms between depth frames Colormap Show RGB underlay Underlay opacity

Timing

Last infer: — · avg (20): — · FPS: —
Model load: — · Tensor: —