ohaiibuzzle/DA3-ONNX

Depth Anything 3, ONNX (static 504×378)

ONNX exports of Depth Anything 3 Mono Large and Depth Anything 3 Small (ByteDance Seed, Apache-2.0), made for ReViBlur.

file	size	sha256	notes
da3mono_l_504x378.onnx	1,336,957,133 B	9b1b3f1495a9273adb4abcdc3e716416a28716b6bda4f13a29da462200e931ec	fp32, weights unchanged. Best; for Snapdragon NPU / GPU
da3mono_l_504x378_int8.onnx	431,206,981 B	416ba24136fb69e2a11298fdae1ecab78e1468bd3982386551027d4048c5d2b9	int8 MatMul weights (dynamic quantization), for CPU / NNAPI
da3_s_504x378_int8.onnx	37,027,601 B	9194a0d5602d71e6407aabeb50ebe5bde9fb25d0c35514b7ef868219034e14d7	DA3-Small, int8; for weak phones. No sky output

All opset 17. The int8 files quantize only MatMul/Gemm weights (onnxruntime quantize_dynamic, QInt8); convolutions stay fp32. They don't run on the QNN HTP (NPU) backend; use QNN GPU, NNAPI or CPU.

Inputs / outputs

name	shape	meaning
image (in)	1×3×378×504 float32	RGB in [0,1], ImageNet-normalised (mean 0.485/0.456/0.406, std 0.229/0.224/0.225), landscape
depth (out)	1×378×504 float32	relative depth, far = large (take 1/depth for disparity)
sky (out, Large only)	1×378×504 float32	sky score, ≥ 0.3 = sky

Static shapes so the Qualcomm QNN HTP (NPU) backend can compile them. DA3's own sky clamp (random sampling + quantile) is left out of the graph; apply it yourself if needed. For portrait-orientation photos, rotate to landscape before inference.

Quality

Spearman rank correlation of 1/depth against a vivo X300 Ultra's own portrait depth, on one photo, with Mac CPU time per run:

file	Spearman	CPU time
Large fp32	0.972	2.4 s
Large int8	0.965	1.4 s
Small int8	0.907	0.18 s

The fp32 exports match PyTorch to a max relative error of 9e-5 (Large) and 1e-5 (Small). DA3-Small's ray head is dropped; that part of its checkpoint is partly untrained, and depth doesn't use it.

How they were made

git clone https://github.com/ByteDance-Seed/Depth-Anything-3
PYTHONPATH=Depth-Anything-3/src python tools/export_da3.py out/           # Large + Large int8
PYTHONPATH=Depth-Anything-3/src python tools/export_da3.py --small out/   # Small + Small int8

tools/ is from the ReViBlur repo.

Licence and attribution

Apache-2.0, same as the original models; see LICENSE. Depth Anything 3 © 2025 ByteDance Ltd. and/or its affiliates. Paper: Depth Anything 3: Recovering the Visual Space from Any Views (Lin et al., 2025).