Introducing Depth Anything for The Foundry's Nuke. This project brings state-of-the-art depth map generation to your favorite compositing software.

Depth Anything (V2) is a neural network that creates accurate depth maps from single images, handling a wide range of subjects.

This tool is natively integrated within Nuke, providing a seamless and streamlined experience. It requires no external dependencies or complicated setup, enabling artists to leverage the cutting-edge AI model in their familiar workflow.

• Simple interface: No 3D tracking setup required
• High detail: Captures relatively difficult edges
• Efficient memory usage: Downrez option allows good results on most 6GB graphics cards
• CPU compatible: Can be batched to CPU render farms

You may also be interested in my other Nuke Cattery nodes or gizmos. Feel free to check them out:

• RIFE for Nuke: Advanced AI-driven retiming and optical flow
• SegmentAnything for Nuke: Cutting-edge AI object segmentation
• VITMatte for Nuke: AI-powered natural edge detail extraction

• Guided Blur/Refine Edge: High-speed edge detail extraction using BlinkScript

Nuke 13.2v9+, tested on Linux and Windows.

1. Download and unzip the latest release from here.
2. Copy the extracted Cattery folder to .nuke or your plugins path.
3. In the toolbar, choose Cattery > Update or simply restart Nuke.

Depth Anything V2 will then be accessible under the toolbar at Cattery > Depth Estimation > DepthAnythingV2.

4. Add the path for *Depth Anything V2 to your init.py:

import nuke
nuke.pluginAddPath('./Cattery/DepthAnythingV2')

5. Add an menu item to the toolbar in your menu.py:

import nuke
toolbar = nuke.menu("Nodes")
toolbar.addCommand('Cattery/Depth Estimation/DepthAnythingV2', 'nuke.createNode("DepthAnythingV2")', icon="DepthAnythingV2.png")

Connect an input image to the Depth Anything node. If necessary, adjust the Near and Far depth values to scale the Z-depth values accordingly.

• Use GPU if available: Utilize GPU acceleration when possible.

• Bypass sRGB Conversion: Skips sRGB color space conversion for input images.

• View: Selects the output display mode:

  • Preview (False Color): Displays the depth map in false colors for easy visualization.
  • Final Output: Output the Z-depth channel into the original input.

• Downrez: Reduces the resolution of the input image to optimize memory usage and processing speed.

• Far: Sets the maximum depth value in the scene, representing the farthest objects.

• Near: Defines the minimum depth value, corresponding to the closest objects in the scene.

• Invert Map: Reverses the depth values, making near objects appear far and vice versa.

The current release includes the V2_Small model, which is the best commercially allowed model according to the original project's licensing terms. However, more capable models can be converted to .cat format by following the compilation steps below.

Tests under Rocky Linux 8, AMD Ryzen Threadripper 3960X 24-Core, NVidia GeForce RTX 3090.

To retrain or modify the model for use with Nuke's CatFileCreator, you'll need to convert it into the PyTorch format .pt. Below are the primary methods to achieve this:

Google Colaboratory offers a free, cloud-based development environment ideal for experimentation or quick modifications. It's important to note that Colaboratory uses Python 3.10, which is incompatible with the PyTorch version (1.6) required by Nuke 13.

For those targetting Nuke 14 or 15, Google Colaboratory is a convenient choice.

Compiling the model locally gives you full control over the versions of Python, PyTorch, and CUDA you use. Setting up older versions, however, can be challenging.

For Nuke 13, which requires PyTorch 1.6, using Docker is highly recommended. This recommendation stems from the lack of official PyTorch package support for CUDA 11.

Fortunately, Nvidia offers Docker images tailored for various GPUs. The Docker image version 20.07 is specifically suited for PyTorch 1.6.0 + CUDA 11 requirements.

Access to these images requires registration on Nvidia's NGC Catalog.

Once Docker is installed on your system, execute the following command to initiate a terminal within the required environment. You can then clone the repository and run python nuke_dan.py to compile the model.

docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:20.07-py3
git clone https://github.com/rafaelperez/Depth-Anything-for-Nuke.git
cd Depth-Anything-for-Nuke

# Compiles the default model, V2 small, half precision.
python nuke_dan.py

# Compiles the V2 large model, half precision.
python nuke_dan.py --version v2 --model-size vitl --half

# Check the available models and options with
python nuke_dan.py --help

For projects targeting Nuke 14+, which requires PyTorch 1.12, you can use the following Docker image, version 22.05:

docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:22.05-py3

For more information on selecting the appropriate Python, PyTorch, and CUDA combination, refer to Nvidia's Framework Containers Support Matrix.

This repository, DepthAnythingV2.cat is licensed under the MIT License, and is derived from https://github.com/LiheYoung/Depth-Anything and https://github.com/DepthAnything/Depth-Anything-V2.

While the MIT License permits commercial use of Depth Anything, the dataset used for its training may be under a non-commercial license.

This license does not cover the underlying pre-trained model, associated training data, and dependencies, which may be subject to further usage restrictions.

Always refer to https://github.com/LiheYoung/Depth-Anything and https://github.com/DepthAnything/Depth-Anything-V2 for the most up-to-date information on associated licensing terms.

If you find this project useful, please consider citing:

@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}
}

@inproceedings{depth_anything_v1,
  title={Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data}, 
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  booktitle={CVPR},
  year={2024}
}