CVPR 2024 (Oral, Best Paper Award Candidate)

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.

2024-05-28: Training code is released.
2024-03-23: Added LCM v1.0 for faster inference - try it out at
2024-03-04: Accepted to CVPR 2024.
2023-12-22: Contributed to Diffusers community pipeline.
2023-12-19: Updated license to Apache License, Version 2.0.
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).

We offer several ways to interact with Marigold:

1. We integrated Marigold Pipelines into diffusers 🧨. Check out many exciting usage scenarios in this diffusers tutorial.

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

3. Run the demo locally (requires a GPU and an nvidia-docker2, see Installation Guide):

   1. Paper version: docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/toshas-marigold:latest python app.py
   2. LCM version: docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/prs-eth-marigold-lcm:latest python app.py

4. Extended demo on a Google Colab:

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

6. Finally, local development instructions with this codebase are given below.

The inference code was tested on:

• Ubuntu 22.04 LTS, Python 3.10.12, CUDA 11.7, GeForce RTX 3090 (pip)

We recommend running the code in WSL2:

1. Install WSL following installation guide.
2. Install CUDA support for WSL following installation guide.
3. Find your drives in /mnt/<drive letter>/; check WSL FAQ for more details. Navigate to the working directory of choice.

Clone the repository (requires git):

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

We provide several ways to install the dependencies.

1. Using Mamba, which can installed together with Miniforge3.

   Windows users: Install the Linux version into the WSL.

   After the installation, Miniforge needs to be activated first: source /home/$USER/miniforge3/bin/activate.

   Create the environment and install dependencies into it:

   mamba env create -n marigold --file environment.yaml
   conda activate marigold

2. Using pip: Alternatively, create a Python native virtual environment and install dependencies into it:

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

Keep the environment activated before running the inference script. Activate the environment again after restarting the terminal session.

1. Use selected images from our paper:

   bash script/download_sample_data.sh

2. Or place your images in a directory, for example, under input/in-the-wild_example, and run the following inference command.

The LCM checkpoint is distilled from our original checkpoint towards faster inference speed (by reducing inference steps). The inference steps can be as few as 1 (default) to 4. Run with default LCM setting:

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

This setting corresponds to our paper. For academic comparison, please run with this setting.

python run.py \
    --checkpoint prs-eth/marigold-v1-0 \
    --denoise_steps 50 \
    --ensemble_size 10 \
    --input_rgb_dir input/in-the-wild_example \
    --output_dir output/in-the-wild_example

You can find all results in output/in-the-wild_example. Enjoy!

The default settings are optimized for the best result. However, the behavior of the code can be customized:

• Trade-offs between the accuracy and speed (for both options, larger values result in better accuracy at the cost of slower inference.)

  • --ensemble_size: Number of inference passes in the ensemble. For LCM ensemble_size is more important than denoise_steps. Default: 10 5 (for LCM).
  • --denoise_steps: Number of denoising steps of each inference pass. For the original (DDIM) version, it's recommended to use 10-50 steps, while for LCM 1-4 steps. When unassigned (None), will read default setting from model config. Default: 10 4 (for LCM) None.

• By default, the inference script resizes input images to the processing resolution, and then resizes the prediction back to the original resolution. This gives the best quality, as Stable Diffusion, from which Marigold is derived, performs best at 768x768 resolution.

  • --processing_res: the processing resolution; set as 0 to process the input resolution directly. When unassigned (None), will read default setting from model config. Default: 768 None.
  • --output_processing_res: produce output at the processing resolution instead of upsampling it to the input resolution. Default: False.
  • --resample_method: the resampling method used to resize images and depth predictions. This can be one of bilinear, bicubic, or nearest. Default: bilinear.

• --half_precision or --fp16: Run with half-precision (16-bit float) to have faster speed and reduced VRAM usage, but might lead to suboptimal results.

• --seed: Random seed can be set to ensure additional reproducibility. Default: None (unseeded). Note: forcing --batch_size 1 helps to increase reproducibility. To ensure full reproducibility, deterministic mode needs to be used.

• --batch_size: Batch size of repeated inference. Default: 0 (best value determined automatically).

• --color_map: Colormap used to colorize the depth prediction. Default: Spectral. Set to None to skip colored depth map generation.

• --apple_silicon: Use Apple Silicon MPS acceleration.

By default, the checkpoint is stored in the Hugging Face cache. The HF_HOME environment variable defines its location and can be overridden, e.g.:

export HF_HOME=$(pwd)/cache

Alternatively, use the following script to download the checkpoint weights locally:

bash script/download_weights.sh marigold-v1-0
# or LCM checkpoint
bash script/download_weights.sh marigold-lcm-v1-0

At inference, specify the checkpoint path:

python run.py \
    --checkpoint checkpoint/marigold-v1-0 \
    --denoise_steps 50 \
    --ensemble_size 10 \
    --input_rgb_dir input/in-the-wild_example\
    --output_dir output/in-the-wild_example

Install additional dependencies:

pip install -r requirements+.txt -r requirements.txt

Set data directory variable (also needed in evaluation scripts) and download evaluation datasets into corresponding subfolders:

export BASE_DATA_DIR=<YOUR_DATA_DIR>  # Set target data directory

wget -r -np -nH --cut-dirs=4 -R "index.html*" -P ${BASE_DATA_DIR} https://share.phys.ethz.ch/~pf/bingkedata/marigold/evaluation_dataset/

Run inference and evaluation scripts, for example:

# Run inference
bash script/eval/11_infer_nyu.sh

# Evaluate predictions
bash script/eval/12_eval_nyu.sh

Note: although the seed has been set, the results might still be slightly different on different hardware.

Based on the previously created environment, install extended requirements:

pip install -r requirements++.txt -r requirements+.txt -r requirements.txt

Set environment parameters for the data directory:

export BASE_DATA_DIR=YOUR_DATA_DIR  # directory of training data
export BASE_CKPT_DIR=YOUR_CHECKPOINT_DIR  # directory of pretrained checkpoint

Download Stable Diffusion v2 checkpoint into ${BASE_CKPT_DIR}

Prepare for Hypersim and Virtual KITTI 2 datasets and save into ${BASE_DATA_DIR}. Please refer to this README for Hypersim preprocessing.

Run training script

python train.py --config config/train_marigold.yaml

Resume from a checkpoint, e.g.

python train.py --resume_run output/marigold_base/checkpoint/latest

Evaluating results

Only the U-Net is updated and saved during training. To use the inference pipeline with your training result, replace unet folder in Marigold checkpoints with that in the checkpoint output folder. Then refer to this section for evaluation.

Note: Although random seeds have been set, the training result might be slightly different on different hardwares. It's recommended to train without interruption.

Please refer to this instruction.

Please cite our paper:

@InProceedings{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},
      booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2024}
}

This work is licensed under the Apache License, Version 2.0 (as defined in the LICENSE).

By downloading and using the code and model you agree to the terms in the LICENSE.