Fine-Tuning Image-Conditional Diffusion Models

[Paper] [Project Page] [HF demo depth] [HF demo normals] [BibTeX]

📢 News

• 2024-10-28: Accepted to WACV 2025.
• 2024-10-23: Training code release.
• 2024-09-24: Evaluation code release.
• 2024-09-18: Inference code release.

⏩ Quickstart

pip install torch diffusers transformers accelerate

from diffusers import DiffusionPipeline
import diffusers

image = diffusers.utils.load_image(
    "https://gonzalomartingarcia.github.io/diffusion-e2e-ft/static/lego.jpg"
)

# Depth
pipe = DiffusionPipeline.from_pretrained(
    "GonzaloMG/marigold-e2e-ft-depth",
    custom_pipeline="GonzaloMG/marigold-e2e-ft-depth",
).to("cuda")
depth = pipe(image)
pipe.image_processor.visualize_depth(depth.prediction)[0].save("depth.png")
pipe.image_processor.export_depth_to_16bit_png(depth.prediction)[0].save("depth_16bit.png")


# Normals
pipe = DiffusionPipeline.from_pretrained(
    "GonzaloMG/stable-diffusion-e2e-ft-normals",
    custom_pipeline="GonzaloMG/marigold-e2e-ft-normals",
).to("cuda")
normals = pipe(image)
pipe.image_processor.visualize_normals(normals.prediction)[0].save("normals.png")

🔧 Development Setup

Tested with Python 3.10.

1. Clone repository:

git clone https://github.com/VisualComputingInstitute/diffusion-e2e-ft.git
cd diffusion-e2e-ft

2. Install dependencies:

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

🤖 Models

The following checkpoints are available for inference. Note that the Marigold (Depth) and GeoWizard (Depth & Normals) diffusion estimators are the official checkpoints provided by their respective authors and were not trained by us. Following the Marigold training regimen, we have trained a Marigold diffusion estimator for normals.

"E2E FT" denotes models we have fine-tuned end-to-end on task-specific losses, either starting from the pretrained diffusion estimator or directly from Stable Diffusion. Since the fine-tuned models are single-step deterministic models, the noise should always be zeros and the ensemble size and number of inference steps should always be 1.

Models	Diffusion Estimator	Stable Diffusion + E2E FT	Diffusion Estimator + E2E FT
Marigold (Depth)	prs-eth/marigold-depth-v1-0	GonzaloMG/stable-diffusion-e2e-ft-depth	GonzaloMG/marigold-e2e-ft-depth
Marigold (Normals)	GonzaloMG/marigold-normals	GonzaloMG/stable-diffusion-e2e-ft-normals	GonzaloMG/marigold-e2e-ft-normals
GeoWizard (Depth&Normals)	lemonaddie/geowizard	N/A	GonzaloMG/geowizard-e2e-ft

🏃 Inference

1. Marigold checkpoints:

python Marigold/run.py \
    --checkpoint="GonzaloMG/marigold-e2e-ft-depth" \
    --modality depth \
    --input_rgb_dir="input" \
    --output_dir="output/marigold_ft"

python Marigold/run.py \
    --checkpoint="GonzaloMG/marigold-e2e-ft-normals" \
    --modality normals \
    --input_rgb_dir="input" \
    --output_dir="output/marigold_ft"

Argument	Description
--checkpoint	Hugging Face model path.
--modality	Output modality; depth or normals.
--input_rgb_dir	Path to the input images.
--output_dir	Path to the output depth or normal images.
--denoise_steps	Number of inference steps; default 1 for E2E FT models.
--ensemble_size	Number of samples for ensemble; default 1 for E2E FT models.
--timestep_spacing	Defines how timesteps are distributed; trailing or leading; default trailing for the fixed inference schedule.
--noise	Noise types; gaussian, pyramid, or zeros; default zeros for E2E FT models.
--processing_res	Resolution the model uses for generation; 0 for matching the RGB input resolution; default 768.
--output_processing_res	If True, the generated image is not resized to match the RGB input resolution; default False.
--half_precision	If True, operations are performed in half precision; default False.
--seed	Sets the seed.
--batch_size	Batched inference when ensembling; default 1.
--resample_method	Resampling method used for resizing the RGB input and generated output; bilinear, bicubic, or nearest; default bilinear.

2. GeoWizard checkpoints:

python GeoWizard/run_infer.py \
    --pretrained_model_path="GonzaloMG/geowizard-e2e-ft" \
    --domain indoor \
    --input_dir="input" \
    --output_dir="output/geowizard_ft"

Argument	Description
--pretrained_model_path	Hugging Face model path.
--domain	Domain with respect to the RGB input; indoor, outdoor, or object.
--input_dir	Path to the input images.
--output_dir	Path to the output depth and normal images.
--denoise_steps	Number of inference steps; default 1 for E2E FT models.
--ensemble_size	Number of samples for ensemble; default 1 for E2E FT models.
--timestep_spacing	Defines how timesteps are distributed; trailing or leading; default trailing for the fixed inference schedule.
--noise	Noise types; gaussian, pyramid, or zeros; default zeros for E2E FT models.
--processing_res	Resolution the model uses for generation; 0 for matching the RGB input resolution; default 768.
--output_processing_res	If True, the generated image is not resized to match the RGB input resolution; default False.
--half_precision	If True, operations are performed in half precision; default False.
--seed	Sets the seed.

By using the correct trailing timestep spacing, it is possible to sample single to few-step depth maps and surface normals from diffusion estimators. These samples will be blurry but become sharper by increasing the number of inference steps, e.g., from 10 to 50. Metrics can be improved by increasing the ensemble size, e.g., to 10. Since diffusion estimators are probabilistic models, the noise setting can be adjusted to either gaussian noise or multiresolution pyramid noise.

Our single-step deterministic E2E FT models outperform the previously mentioned diffusion estimators.

📋 Metrics

Depth Method	Inference Time	NYUv2 AbsRel↓	KITTI AbsRel↓	ETH3D AbsRel↓	ScanNet AbsRel↓	DIODE AbsRel↓
Stable Diffusion + E2E FT	121ms	5.4	9.6	6.4	5.8	30.3
Marigold + E2E FT	121ms	5.2	9.6	6.2	5.8	30.2
GeoWizard + E2E FT	254ms	5.6	9.8	6.3	5.9	30.6

Normals Method	Inference Time	NYUv2 Mean↓	ScanNet Mean↓	iBims-1 Mean↓	Sintel Mean↓
Stable Diffusion + E2E FT	121ms	16.5	15.3	16.1	33.5
Marigold + E2E FT	121ms	16.2	14.7	15.8	33.5
GeoWizard + E2E FT	254ms	16.1	14.7	16.2	33.4

Inference time is for a single 576x768-pixel image, evaluated on an NVIDIA RTX 4090 GPU.

📊 Evaluation

We utilize the official Marigold evaluation pipeline to evaluate the affine-invariant depth estimation checkpoints, and we use the official DSINE evaluation pipeline to evaluate the surface normals estimation checkpoints. The code has been streamlined to exclude unnecessary parts, and changes have been marked.

Depth

The Marigold evaluation datasets can be downloaded to data/marigold_eval/ at the root of the project using the following snippet:

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

After downloading, the folder structure should look as follows:

data
└── marigold_eval
    ├── diode
    │   └── diode_val.tar
    ├── eth3d
    │   └── eth3d.tar
    ├── kitti
    │   └── kitti_eigen_split_test.tar
    ├── nyuv2
    │   └── nyu_labeled_extracted.tar
    └── scannet
        └── scannet_val_sampled_800_1.tar

Run the 0_infer_eval_all.sh script to evaluate the desired model on all datasets.

./experiments/depth/eval_args/marigold_e2e_ft/0_infer_eval_all.sh

./experiments/depth/eval_args/stable_diffusion_e2e_ft/0_infer_eval_all.sh

./experiments/depth/eval_args/geowizard_e2e_ft/0_infer_eval_all.sh

The evaluation results for the selected model are located in the experiments/depth/marigold directory. For a given dataset, the script first performs the necessary inference, storing the estimations in a prediction folder. Later, these depth maps are aligned and evaluated against the ground truth. Metrics and evaluation settings are available as .txt files.

<model>
└── <dataset>
    ├── arguments.txt
    ├── eval_metric
    │   └── eval_metrics-least_square.txt
    └── prediction

Normals

The DSINE evaluation datasets (dsine_eval.zip) should be extracted into the data folder at the root of the project. The folder structure should look as follows:

data
└── dsine_eval
   ├── ibims
   ├── nyuv2
   ├── oasis
   ├── scannet
   ├── sintel
   └── vkitti

Run the following commands to evaluate the models on all datasets.

python -m DSINE.projects.dsine.test \
    experiments/normals/eval_args/marigold_e2e_ft.txt \
    --mode benchmark

python -m DSINE.projects.dsine.test \
    experiments/normals/eval_args/stable_diffusion_e2e_ft.txt \
    --mode benchmark

python -m DSINE.projects.dsine.test \
    experiments/normals/eval_args/geowizard_e2e_ft.txt \
    --mode benchmark

Evaluation results are saved in the experiments/normals/dsine folder. This includes the used settings (params.txt) and the metrics for each <dataset> (metrics.txt).

dsine
  └── <model-type/model>
      ├── log
      │   └── params.txt
      └── test
          └── <dataset>
              └── metrics.txt

🏋️ Training

Datasets

The fine-tuned models are trained on the Hypersim and Virtual KITTI 2 datasets.

Hypersim

Download the Hypersim dataset using the dataset_download_images.py script and unzip the files to data/hypersim/raw_data at the root of the project. Download the scene split file from the Hypersim repository and place it in data/hypersim.

data
└── hypersim
    ├── metadata_images_split_scene_v1.csv
    └── raw_data
        ├── ai_001_001
        ├── ...
        └── ai_055_010

Run Marigold's preprocessing script, which will save the processed data to data/hypersim/processed.

python Marigold/script/dataset_preprocess/hypersim/preprocess_hypersim.py \
  --split_csv data/hypersim/metadata_images_split_scene_v1.csv

Download the surface normals in png format using Hypersim's download.py script.

./download.py --contains normal_cam.png --silent

Place the downloaded surface normals in data/hypersim/processed/normals.

The final processed file structure should look like this:

data
└── hypersim
    └── processed
        ├── normals
        │   ├── ai_001_001
        │   ├── ...
        │   └── ai_055_010
        └── train
            ├── ai_001_001
            ├── ...
            ├── ai_055_010
            └── filename_meta_train.csv

Virtual KITTI 2

Download the RGB (vkitti_2.0.3_rgb.tar) and depth (vkitti_2.0.3_depth.tar) files from the official website. Place them in data/virtual_kitti_2 at the root of the project and finally extract them using the following shell commands.

mkdir vkitti_2.0.3_rgb && tar -xf vkitti_2.0.3_rgb.tar -C vkitti_2.0.3_rgb
mkdir vkitti_2.0.3_depth && tar -xf vkitti_2.0.3_depth.tar -C vkitti_2.0.3_depth

Virtual KITTI 2 does not provide surface normals. Therefore, we estimate them from the depth maps using discontinuity-aware gradient filters. Run our provided script to generate the normals which will be saved to data/virtual_kitti_2/vkitti_DAG_normals.

python depth-to-normal-translator/python/gen_vkitti_normals.py

The final processed file structure should look like this:

data
└── virtual_kitti_2
    ├── vkitti_2.0.3_depth
    │   ├── Scene01
    │   ├── Scene02
    │   ├── Scene06
    │   ├── Scene18
    │   └── Scene20
    ├── vkitti_2.0.3_rgb
    │   ├── Scene01
    │   ├── Scene02
    │   ├── Scene06
    │   ├── Scene18
    │   └── Scene20
    └── vkitti_DAG_normals
        ├── Scene01
        ├── Scene02
        ├── Scene06
        ├── Scene18
        └── Scene20

E2E FT Model Training

To train the end-to-end fine-tuned depth and normals models, run the scripts in the training/scripts directory:

./training/scripts/train_marigold_e2e_ft_depth.sh

./training/scripts/train_stable_diffusion_e2e_ft_depth.sh

./training/scripts/train_marigold_e2e_ft_normals.sh

./training/scripts/train_stable_diffusion_e2e_ft_normals.sh

./training/scripts/train_geowizard_e2e_ft.sh

The fine-tuned models will be saved to model-finetuned at the root of the project.

model-finetuned
    └── <model>
        ├── arguments.txt
        ├── model_index.json
        ├── text_encoder # or image_encoder for GeoWizard
        ├── tokenizer
        ├── feature_extractor
        ├── scheduler
        ├── vae
        └── unet 

Note

For multi GPU training, set the desired number of devices and nodes in the training/scripts/multi_gpu.yaml file and replace accelerate launch with accelerate launch --multi_gpu --config_file training/scripts/multi_gpu.yaml in the training scripts.

🎓 Citation

If you use our work in your research, please use the following BibTeX entry.

@InProceedings{martingarcia2024diffusione2eft,
  title     = {Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think},
  author    = {Martin Garcia, Gonzalo and Abou Zeid, Karim and Schmidt, Christian and de Geus, Daan and Hermans, Alexander and Leibe, Bastian},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2025}
}