Depth Anything V3 (depth_da3)

Per-frame · Monocular depth estimation

Estimate a per-pixel depth map from a single RGB frame.

What you give it

• An RGB clip.

What you get back

• A per-frame depth map encoded as image data (relative depth by default; metric depth with the DA3METRIC-LARGE variant).

Typical VFX uses: depth-based defocus, atmospheric haze, depth-driven roto assistance, parallax 2.5D moves, relighting passes — without needing tracked geometry or stereo capture.

Commercial use

Variant	License	Commercial OK?
DA3-SMALL	Apache 2.0	✅ Yes
DA3-BASE	Apache 2.0	✅ Yes
DA3METRIC-LARGE	Apache 2.0	✅ Yes
DA3MONO-LARGE	Apache 2.0	✅ Yes
DA3-LARGE	CC BY-NC 4.0	❌ No
DA3-GIANT	CC BY-NC 4.0	❌ No
DA3NESTED-*	CC BY-NC 4.0	❌ No

For paid production work, use one of the Apache-licensed variants.

Requirements

• GPU VRAM (approximate):
  • SMALL: ~2 GB
  • BASE: ~3 GB
  • LARGE: ~6 GB
  • GIANT: ~12 GB+
• ComfyUI custom node: PozzettiAndrea/ComfyUI-DepthAnythingV3 (MIT)
• Model weights: the embedded workflow’s DownloadAndLoadDepthAnythingV3Model node auto-downloads the selected model variant to ComfyUI/models/depthanything3/ on first use. Hugging Face source: depth-anything/DA3-*.

Parameters

The plugin exposes the standard ComfyUI base parameters (server URL, mount paths, project name, workflow path) plus:

Parameter	Meaning
Normalization	How depth is normalized. V2-Style (default) gives normalized 0–1 depth with sky-mask support, recommended for ControlNet; Raw gives unnormalized metric depth for 3D reconstruction.
Resize Method	How the input is fitted to the model’s resolution: Resize (scale to fit, default), Pad (preserve aspect ratio), or Crop (center-crop).
Invert Depth	Swap near and far, so near becomes far and far becomes near. Off by default.
Keep Model Size	Output depth at the model’s internal resolution instead of upscaling back to the input resolution. Off by default.
Model Variant	Which DA3 checkpoint to load. Options: DA3-Small (80M, fast), DA3-Base (220M, balanced), DA3-Large (350M, high quality — default), DA3-Giant (1.15B, best quality), DA3-Mono-Large and DA3-Metric-Large (350M, sky-mask support), DA3-Nested-Giant-Large (1.4B, combined model with metric scaling).
Precision	Computation precision: Auto (default, picks the best for your GPU), FP16 (faster, less VRAM), FP32 (slower, more accurate), or BF16 (Ampere+ GPUs).
Attention	Attention implementation: Flash Attention (fastest, default; needs the flash-attn library), xFormers (fast alternative), or Math (standard PyTorch, always available).

Demos & comparisons

Input → output

A DA3 depth map, which is the only thing this plugin returns. Crop of the “Depth & Ray Maps” panel of the project-page teaser. © Lin et al., ByteDance Seed, 2025; reproduced for documentation under Apache 2.0 attribution. Source: depth-anything-3.github.io.

A note on the upstream demos. DA3’s own promotional material — the teaser reel, the project-page gallery, the GitHub README — leads with multi-view geometry: point clouds, camera trajectories, 3D reconstruction from unposed images. This plugin ships none of that. Its workflow is LoadEXR → DepthAnything_V3 → SaveEXR: one frame in, one depth map out. The reconstruction path would need a different ComfyUI graph, and a point cloud is not something an OFX image effect can hand back to the host. Read the demos below with that in mind:

• Project page — depth-anything-3.github.io — gallery and comparisons against DA2 and VGGT.
• Hugging Face Space — interactive demo — upload your own image and see the result.
• GitHub README — ByteDance-Seed/Depth-Anything-3 — side-by-side RGB / depth / point-cloud comparisons.

Image attribution: ByteDance Seed; reproduced for documentation purposes with citation to arXiv:2511.10647. See the credits page.

Limitations

• Depth is relative / affine-invariant unless you use the metric variant. Do not interpret raw values as world-scale distances.
• Transparent surfaces (glass, water, smoke) collapse depth ambiguously.
• Mirrors and strong specular highlights return the depth of the reflected scene, not the surface.
• Heavy motion blur degrades stability. The model has no temporal consistency for video — consider DepthCrafter when temporal stability matters more than per-frame fidelity.

Credits

This plugin is a thin wrapper around the work of:

Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, Bingyi Kang. Depth Anything 3: Recovering the Visual Space from Any Views. arXiv preprint arXiv:2511.10647, 2025. ByteDance Seed. Paper · Project page · GitHub

ComfyUI node by PozzettiAndrea.

Citation

@article{depthanything3,
  title   = {Depth Anything 3: Recovering the Visual Space from Any Views},
  author  = {Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and
             Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
  journal = {arXiv preprint arXiv:2511.10647},
  year    = {2025}
}