Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation

This repository represents the official implementation of the paper titled "Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation".

Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, Konrad Schindler

We present Marigold, a diffusion model and associated fine-tuning protocol for monocular depth estimation. Its core principle is to leverage the rich visual knowledge stored in modern generative image models. Our model, derived from Stable Diffusion and fine-tuned with synthetic data, can zero-shot transfer to unseen data, offering state-of-the-art monocular depth estimation results.

📢 News

2023-12-08: Added - try it out with your images for free!
2023-12-05: Added - dive deeper into our inference pipeline!
2023-12-04: Added paper and inference code (this repository).

Usage

We offer a number of way to interact with Marigold:

1. A free online interactive demo is available here: (kudos to the HF team for the GPU grant)

2. Run the demo locally (requires a GPU and an nvidia-docker2, see Installation Guide): docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/toshas-marigold:latest python app.py

3. Extended demo on a Google Colab:

4. If you just want to just see the examples, visit our gallery:

5. Finally, local development instructions are given below.

🛠️ Setup

This code has been tested on:

• Python 3.10.12, PyTorch 2.0.1, CUDA 11.7, GeForce RTX 3090
• Python 3.10.4, Pytorch 2.0.1, CUDA 11.7, GeForce RTX 4090

📦 Repository

git clone https://github.com/prs-eth/Marigold.git
cd Marigold

💻 Dependencies

python -m venv venv/marigold
source venv/marigold/bin/activate
pip install -r requirements.txt

🚀 Inference on in-the-wild images

📷 Sample images

bash script/download_sample_data.sh

🎮 Inference

This script will automatically download the checkpoint.

python run.py \
    --input_rgb_dir data/in-the-wild_example \
    --output_dir output/in-the-wild_example

⚙️ Inference settings

• The inference script by default will resize the input images and resize back to the original resolution.

  • --resize_to_max_res: The maximum edge length of resized input image. Default: 768.
  • --not_resize_input: If given, will not resize the input image.
  • --not_resize_output: If given, will not resize the output image back to the original resolution. Only valid without --not_resize_input option.

• Trade-offs between accuracy and speed (for both options, larger value results in more accurate results at the cost of slower inference speed.)

  • --n_infer: Number of inference passes to be ensembled. Default: 10.
  • --denoise_steps: Number of diffusion denoising steps of each inference pass. Default: 10.

• --seed: Random seed, can be set to ensure reproducibility. Default: None (using current time as random seed).

• --depth_cmap: Colormap used to colorize the depth prediction. Default: Spectral.

• The model cache directory can be controlled by environment variable HF_HOME, for example:

  export HF_HOME=$(pwd)/checkpoint

⬇ Using local checkpoint

# Download checkpoint
bash script/download_weights.sh

python run.py \
    --checkpoint checkpoint/Marigold_v1_merged \
    --input_rgb_dir data/in-the-wild_example\
    --output_dir output/in-the-wild_example

🎓 Citation

@misc{ke2023repurposing,
      title={Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation}, 
      author={Bingxin Ke and Anton Obukhov and Shengyu Huang and Nando Metzger and Rodrigo Caye Daudt and Konrad Schindler},
      year={2023},
      eprint={2312.02145},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.