Depth Anything V2

• Original code: https://github.com/DepthAnything/Depth-Anything-V2 at a561b84
• Code licence: Apache-2.0 (permissive); files: LICENSE
• Weights licence: Apache-2.0
  • depth-anything-v2-b: CC-BY-NC-4.0
  • depth-anything-v2-l: CC-BY-NC-4.0
• Licence evidence: https://github.com/DepthAnything/Depth-Anything-V2/blob/main/README.md (LICENSE section) and the Hugging Face model cards
• Training data: Teacher (ViT-G) on 595K synthetic images (BlendedMVS, Hypersim, IRS, TartanAir, VKITTI 2); students on 62M pseudo-labelled real images (SA-1B, Open Images, BDD100K, Google Landmarks, ImageNet-21K, LSUN, Objects365, Places365) per the paper
• Licence note: Only the Small weights are Apache-2.0; the Base and Large weights are CC-BY-NC-4.0. The training data of even the Small model has unclear commercial terms (a note only).
• Options: short_side = 518 (>= 14; also input_size)
• Kernel libraries: fastbb, dpt_fused
• Variants: depth-anything-v2-s, depth-anything-v2-b, depth-anything-v2-l
• Outputs: inverse_depth; kind: affine_invariant_inverse
• Input size: any H x W >= 8, resized or padded inside the model (ViT patch 14)
• Tested up to (portable, fp16, RTX 5060 Ti 16 GB): depth-anything-v2-s 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); depth-anything-v2-b 3000x4000 (0.8 GiB peak allocated, 0.9 reserved); depth-anything-v2-l 3000x4000 (1.8 GiB peak allocated, 2.2 reserved)
• depth-anything-v2-s 994x1482, first call: portable + torch.compile 16 s (cold Inductor cache), TensorRT fp16 30 s (ONNX + engine).
• depth-anything-v2-s original 994x1482: torch.autocast(fp16) around the forward
• depth-anything-v2-b original 994x1482: torch.autocast(fp16) around the forward
• depth-anything-v2-l original 994x1482: torch.autocast(fp16) around the forward
• depth-anything-v2-s load / first call: original 0.71 s / 0.38 s, portable 0.11 s / 0.027 s, kernels 0.08† s / 0.009 s
• depth-anything-v2-b load / first call: original 3.94 s / 0.08 s, portable 0.85 s / 0.019 s, kernels 0.2† s / 0.012 s
• depth-anything-v2-l load / first call: original 5.66 s / 0.149 s, portable 2.39 s / 0.081 s, kernels 0.91† s / 0.055 s
• † measured in the speed sweep after the portable load (or an earlier size) in the same process: file cache and kernels are warm, so the time is lower than a fresh start.
• Protocol, median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls

Measured

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the original code named under Details, same checkpoint and protocol.

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10 / 939adfe8, 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.5 % and p99 <= 2 % of the reference's median |value| (absolute error of the output inverse_depth).