◂▸
ZipDepth (ECCV 2026, MIT) is a 6.1M-parameter depth network distilled from Depth Anything V2 Large. We swapped it into our live 2D-to-3D converter and scored it against the same true stereo. Depth takes 3 ms a frame instead of 12, at almost the same quality. In the browser it's a 12 MB download instead of 50 MB.
What we tried
- Loaded ZipDepth's fused inference model in FP16, channels-last, replayed as a CUDA graph like our Depth Anything V2 path.
- Checked its output direction against true depth (0.90 correlation with inverse depth), then ran it through the same stabiliser and stereo warp.
- Exported it to a 12 MB half-precision ONNX file for the browser, run with ONNX Runtime Web on WebGPU.
What we measured
| Measure | Depth Anything V2 Small | ZipDepth | Note |
|---|---|---|---|
| Depth per 720p frame | 11.8 ms | 3.0 ms | |
| 720p throughput | 46.6 fps | 75.2 fps | Median of 3 runs |
| 1080p throughput | 33.5 fps | 38.2 fps | The stereo warp is now the slowest stage |
| Added latency, 720p live | 18.8 ms | 13.8 ms | |
| Depth error vs true stereo | 2.14 px | 2.51 px | |
| Right-eye match (PSNR) | 17.04 dB | 16.93 dB | |
| Browser: depth per frame | 20 ms, 50 MB | 11 ms, 12 MB | Chromium, WebGPU, same RTX 4070 |
What went wrong
- ZipDepth's export script crashed on Windows printing an arrow character; setting UTF-8 output fixed it.
- Converting to half precision changed outputs by up to 4.7% on random input. We kept it, because the smaller download matters more in the browser.
- With depth this cheap, the stereo warp and packing now take most of each frame, so that's the next thing to speed up.
What happens next
- A faster stereo warp: a custom CUDA kernel in place of scatter operations.
- Two live 720p feeds on one card.
- Phones and laptops: measure the browser version on a mid-range phone.
Built with
- ZipDepth base MIT
- Depth Anything V2 Small (comparison) Apache-2.0
- ONNX Runtime Web MIT
- PyTorch BSD-3
Next experiment
Which small local model reads UK paperwork without making numbers up →
Want this for your data?