DA3Mono-LARGE-MLX-fp16

DA3Mono-LARGE-MLX-fp16

depth-anything/DA3MONO-LARGE (Depth Anything 3 Mono, ByteDance) converted to MLX for Apple silicon.

Single-image monocular depth: DinoV2 ViT-L/14 backbone + DPT head with an auxiliary sky-logit branch. Positional embedding is pre-baked (bicubic) for a fixed 504×504 input.

• Input: (1, 3, 504, 504) float32, ImageNet-normalized ((x/255 - mean) / std)
• Outputs: depth (504, 504) = exp(logits), sky (504, 504) = relu(logits)
• Precision: fp16 weights (half the size, ~2x faster; see DA3Mono-LARGE-MLX for fp32)

Use with MLX

import mlx.core as mx
import numpy as np
from PIL import Image
from da3mono_mlx import DA3MonoMLX   # shipped in this repo

model = DA3MonoMLX().load_weights("model.safetensors")

img = Image.open("photo.jpg").convert("RGB").resize((504, 504), Image.BICUBIC)
a = np.asarray(img).astype(np.float32) / 255.0
x = (a - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
x = mx.array(x.transpose(2, 0, 1)[None])          # NCHW

depth, sky = model(x)
mx.eval(depth, sky)

Accuracy vs. PyTorch reference

Validated against the official PyTorch forward on both random tensors and real images:

License & citation

Weights are a conversion of depth-anything/DA3MONO-LARGE; the original model is released under CC-BY-NC-4.0 (non-commercial). Please cite:

@article{depthanything3,
  title={Depth Anything 3: Recovering the Visual Space from Any Views},
  author={Depth Anything Team},
  journal={arXiv preprint arXiv:2511.10647},
  year={2025}
}