Depth Anything 3 for Nuke

Monocular Depth Estimation for Foundry Nuke

Description

This plugin brings Depth Anything 3 monocular depth estimation to Foundry's Nuke compositing software. It generates high-quality depth maps from single RGB images using the DA3Mono-Large model (350M parameters).

Depth Anything 3 represents the state-of-the-art in monocular depth estimation, significantly outperforming previous versions (DA1, DA2) in geometric accuracy.

Pre-traced Models

Pre-traced models for various resolutions (720p, 2K, 3K, 4K) are available on my Patreon.

Usage

⚠️ IMPORTANT: The input resolution MUST match the traced model resolution exactly!

Node Graph Setup

Read → Reformat → Inference → Reformat → Output
       (to model)   (.cat)    (to original)

Reformat Settings (before Inference)

• Type: to box
• Width/Height: exact model resolution
• Resize type: distort
• Filter: Cubic (Keys)
• ☑️ force shape ← critical!

Creating .CAT File

If you need to create a .cat file from the .pt model:

1. Open Nuke's CatFileCreator
2. Set TorchScript file: DepthAnything3_mono_large_WxH_fp32.pt
3. Cat file: DepthAnything3_mono_large_WxH_fp32.cat
4. Channels in: rgba.red, rgba.green, rgba.blue
5. Channels out: rgba.alpha
6. Model ID: DepthAnything3
7. Create the .cat file

For Developers

Requirements for Tracing

• Python 3.10
• PyTorch 2.1.1 (must match Nuke 16.0)
• depth_anything_3 package
• safetensors

Step by Step Installation

conda create -n da3_nuke16 python=3.10 -y
conda activate da3_nuke16
pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
pip install safetensors
git clone https://github.com/ByteDance-Seed/Depth-Anything-3.git
cd Depth-Anything-3
pip install -e .

# Verify
python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA {torch.cuda.is_available()}')"
python -c "from depth_anything_3.api import DepthAnything3; print('DA3 imported OK')"

Download Model

huggingface-cli download depth-anything/DA3MONO-LARGE --local-dir ./DA3MONO-LARGE

Tracing New Resolutions

The included nuke_da3_v3.py script can trace models at any resolution (width and height must be multiples of 14):

# Square resolution
python nuke_da3_v3.py \
    --model-path /path/to/model.safetensors \
    --config-path /path/to/config.json \
    --resolution 2520 \
    --no-half \
    --output-dir ./output

# Non-square (16:9) resolution - saves VRAM!
python nuke_da3_v3.py \
    --model-path /path/to/model.safetensors \
    --config-path /path/to/config.json \
    --width 2058 \
    --height 1092 \
    --no-half \
    --output-dir ./output

Common Resolutions (must be ×14)

Format	Resolution	Patches	VRAM (FP32)
720p	1288×728	92×52	~4 GB
2K DCI	2058×1092	147×78	~6-8 GB
3K DCI	3080×1624	220×116	~12-16 GB
4K DCI	4102×2170	293×155	~20-24 GB

Compatibility

Component	Version
Nuke	15.1, 15.2, 16.0
PyTorch	2.1.1
CUDA	11.8+
OS	Rocky Linux 8/9+, Windows 10/11

Credits

Depth Anything 3 Model

• Paper: Depth Anything 3: Recovering the Visual Space from Any Views
• Authors: Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, Bingyi Kang
• Organization: ByteDance Seed Team
• Repository: https://github.com/ByteDance-Seed/Depth-Anything-3
• Project Page: https://depth-anything-3.github.io/

Nuke Integration

Tracing and wrapper by Peter Mercell

Acknowledgments

Special thanks to:

• Rafael Silva for his excellent Depth-Anything-for-Nuke project, which served as the foundation and inspiration for this work.
• PozzettiAndrea for ComfyUI-DepthAnythingV3 reference implementation.

License

This Nuke integration is released under the Apache License 2.0.

The DA3Mono-Large model is licensed under Apache License 2.0 by ByteDance.

See LICENSE.txt for full license text.

Citation

If you use this in your work, please cite the original paper:

@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}
}