Views
No views yet
(1, 3, 504, 504) float32, ImageNet-normalized ((x/255 - mean) / std)depth (504, 504) = exp(logits), sky (504, 504) = relu(logits)1import mlx.core as mx
2import numpy as np
3from PIL import Image
4from da3mono_mlx import DA3MonoMLX # shipped in this repo
5
6model = DA3MonoMLX().load_weights("model.safetensors")
7
8img = Image.open("photo.jpg").convert("RGB").resize((504, 504), Image.BICUBIC)
9a = np.asarray(img).astype(np.float32) / 255.0
10x = (a - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
11x = mx.array(x.transpose(2, 0, 1)[None]) # NCHW
12
13depth, sky = model(x)
14mx.eval(depth, sky)| metric | value |
|---|---|
| depth Pearson r | 0.9999998 |
| depth median relative error | 0.07 % |
| depth max abs error | 0.02 |
depth-anything/DA3MONO-LARGE; the original
model is released under CC-BY-NC-4.0 (non-commercial). Please cite:1@article{depthanything3,
2 title={Depth Anything 3: Recovering the Visual Space from Any Views},
3 author={Depth Anything Team},
4 journal={arXiv preprint arXiv:2511.10647},
5 year={2025}
6}