ShadeNet-5: albedo, depth and normals from one photo

A 35M-parameter feed-forward network on a frozen DINOv2 backbone. This Space runs the 4/8-bit MIX ONNX export on a free CPU. Depth is relative (per-image scale and shift); normals are x right, y up, z toward the viewer.

Single pass is one forward pass. Single pass + refiner adds a small trained UNet (0.91M parameters) that cleans up the maps, mostly the normals, for a fraction of a second more: it recovers about three quarters of what the slower mode gains. Six passes averages the mirrored and rotated (+-4 degrees) photo and is the best but about six times slower; tick Median to take the median of the six instead of the mean (marginally worse in the card's measurements) and see the difference image. Model card and files. The example photos are Flickr8k photos shown for illustration only; see the model card for their licence status.