open-noodle/distill-any-depth-small

Distill-Any-Depth Small (ONNX), Spatial Scene for Gallery

ONNX export of Distill-Any-Depth Small (arXiv:2502.19204), used by Gallery for Spatial Scene depth maps.

Unmodified weights, exported by machine-learning/depth-model-export/export.py in the Gallery repository. The only graph change is the depth head's final resize, which becomes a scale factor so height and width stay dynamic.

Files

Path	Description
depth/model.onnx	fp32, opset 18
depth/model_int8.onnx	dynamic int8 weights, used on CPU only
config.json	source and preprocessing

Inference contract

Input: pixel_values, shape [1, 3, height, width], float32. RGB scaled to 0..1, then normalised with ImageNet mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225]. Height and width are multiples of 14; Gallery resizes the short side to 518.

Output: predicted_depth, shape [1, height, width], float32. Relative inverse depth (disparity): larger is nearer, with no fixed scale.

Licence

MIT, as the source model.