Depth Anything, fp16 safetensors

Two depth models converted to fp16 safetensors, nothing else changed. They are the pair used by the TS Universal Depth node in comfyui-timesaver: one for single images, one for video.

file	source	what it is for	size
`depth_anything_v2_vitl_fp16.safetensors`	depth-anything/Depth-Anything-V2-Large	a single image	640 MB
`video_depth_anything_vitl_fp16.safetensors`	depth-anything/Video-Depth-Anything-Large	a sequence, temporally consistent	734 MB

⚠️ Licence: CC BY-NC 4.0 — non-commercial

Both upstream checkpoints are published under CC BY-NC 4.0, and this conversion inherits it unchanged. You may not use these weights for commercial purposes. Attribution goes to the Depth Anything authors; this repository holds no rights of its own beyond the conversion.

If you need a permissive licence, the Small variant of Depth Anything V2 is Apache-2.0 — a different model, not a smaller copy of this one.

What was changed

Exactly two things, both mechanical:

float32 → float16. Nothing was pruned, quantised or re-trained.
.pth → .safetensors. The tensors and their names are identical.

Every file carries its provenance in the safetensors metadata: the upstream repository, the upstream filename and the sha256 of the original .pth it was converted from, so the chain back to the source is checkable rather than promised.

from safetensors import safe_open
with safe_open("depth_anything_v2_vitl_fp16.safetensors", framework="pt") as f:
    print(f.metadata())

file	sha256 of the original `.pth`
`depth_anything_v2_vitl_fp16.safetensors`	`a7ea19fa0ed99244e67b624c72b8580b7e9553043245905be58796a608eb9345`
`video_depth_anything_vitl_fp16.safetensors`	`43df27c6b396042ba34ff7b798ab279f64d204d2e86d7a373968f8fa36d0e6fa`

Does fp16 cost quality?

Measured, not assumed. Depth maps were rendered from the same images at 518 px on an RTX 3080 Ti, and each variant compared against pure fp32 with autocast disabled, on the normalised 0..1 depth map:

variant	mean deviation	worst pixel
fp32 weights + autocast	0.020 %	0.39 %
fp16 weights (this repo) + autocast	0.016 – 0.020 %	0.51 %
fp16 weights and fp16 compute	0.021 %	0.61 %

The dominant difference is autocast itself; storing the weights in fp16 adds almost nothing on top of it. Half the download, half the disk, no visible cost.

Loading

from safetensors.torch import load_file
state = load_file("video_depth_anything_vitl_fp16.safetensors")
model.load_state_dict(state)   # architecture from the upstream project

Architectures live in the original projects — Depth-Anything-V2 and Video-Depth-Anything. This repository ships weights only.

Citation

@article{depth_anything_v2,
  title   = {Depth Anything V2},
  author  = {Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen
             and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal = {arXiv:2406.09414},
  year    = {2024}
}

@article{video_depth_anything,
  title   = {Video Depth Anything: Consistent Depth Estimation for Super-Long Videos},
  author  = {Chen, Sili and Guo, Hengkai and Zhu, Shengnan and Zhang, Feihu
             and Huang, Zilong and Feng, Jiashi and Kang, Bingyi},
  journal = {arXiv:2501.12375},
  year    = {2025}
}