Retrieval-Augmented Diffusion Models

arXiv | BibTeX

Retrieval-Augmented Diffusion Models
Andreas Blattmann*, Robin Rombach*, Kaan Oktay, Jonas Müller, Björn Ommer
* equal contribution

News

April 2023

Code release

Requirements

A suitable conda environment named rdm can be created and activated with:

conda env create -f environment.yaml
conda activate rdm

Pretrained Models

A general list of all available checkpoints is available in via our model zoo. If you use any of these models in your work, we are always happy to receive a citation.

Example inference notebooks can be found under scripts/demo_rdm.ipynb and scripts/demo_rarm.ipynb.

Database

To be able to run a RDM/RARM conditioned on a text-prompt and additionally images retrieved from this prompt, you will also need to download the corresponding retrieval database. We provide two distinct databases extracted from the Openimages- and ImageNet datasets. Interchanging the databases results in different capabilities of the model as visualized below, although the learned weights are the same in both cases.

Download the retrieval-databases which contain the retrieval-datasets (Openimages (~18GB) and ImageNet (~1GB)) compressed into CLIP image embeddings:

bash scripts/download_databases.sh

RDM

Since CLIP offers a shared image/text feature space, and RDMs learn to cover a neighborhood of a given example during training, we can directly take a CLIP text embedding of a given prompt and condition on it. Run this mode via

python scripts/rdm_sample.py  --caption "a happy bear reading a newspaper, oil on canvas" --gpu 0

or sample the model unconditionally

python scripts/rdm_sample.py  --gpu 0

RARM

RARMs can be used in a similar manner to RDMs with

python scripts/rarm_sample.py  --caption "a happy bear reading a newspaper, oil on canvas" --gpu 0

and sample the model unconditionally

python scripts/rarm_sample.py  --gpu 0

Train your own RDMs

Data preparation

ImageNet

The code will try to download (through Academic Torrents) and prepare ImageNet the first time it is used. However, since ImageNet is quite large, this requires a lot of disk space and time. If you already have ImageNet on your disk, you can speed things up by putting the data into ${XDG_CACHE}/autoencoders/data/ILSVRC2012_{split}/data/ (which defaults to ~/.cache/autoencoders/data/ILSVRC2012_{split}/data/), where {split} is one of train/validation. It should have the following structure:

${XDG_CACHE}/autoencoders/data/ILSVRC2012_{split}/data/
├── n01440764
│   ├── n01440764_10026.JPEG
│   ├── n01440764_10027.JPEG
│   ├── ...
├── n01443537
│   ├── n01443537_10007.JPEG
│   ├── n01443537_10014.JPEG
│   ├── ...
├── ...

If you haven't extracted the data, you can also place ILSVRC2012_img_train.tar/ILSVRC2012_img_val.tar (or symlinks to them) into ${XDG_CACHE}/autoencoders/data/ILSVRC2012_train/ / ${XDG_CACHE}/autoencoders/data/ILSVRC2012_validation/, which will then be extracted into above structure without downloading it again. Note that this will only happen if neither a folder ${XDG_CACHE}/autoencoders/data/ILSVRC2012_{split}/data/ nor a file ${XDG_CACHE}/autoencoders/data/ILSVRC2012_{split}/.ready exist. Remove them if you want to force running the dataset preparation again.

FFHQ

Create a symlink data/ffhq pointing to the images1024x1024 folder obtained from the FFHQ repository.

Database preparation

Preparation of retrieval database

Download a database of your choice from above.
Precompute nearest neighbors for a given query dataset:
- Create new config for QueryDataset under configs/query_datasets (see template for creation)
- Start nn extraction with python scripts/search_neighbors.py -rc <path to retrieval_config> -qc <path to query config> -s <query dataset split> -bs <batch size> -w <n_workers>, e.e. python scripts/search_neighbors.py --rconfig configs/dataset_builder/openimages.yaml --qc configs/query_datasets/imagenet.yaml -s validation -n

Model Training

Logs and checkpoints for trained models are saved to logs/<START_DATE_AND_TIME>_<config_spec>.

Training autoencoder models

For training autoencoders see the latent diffusion repository.

Training retrieval augmented Models

In configs/rdm/ and configs/rarm/ we provide configs for training RDMs on the FFHQ and ImageNet datasets and RARMs on ImageNet subsets. Training can be started after having downloaded the appropriate files by running

CUDA_VISIBLE_DEVICES=<GPU_ID> python main.py --base configs/<{rdm,rarm}>/<config_spec>.yaml -t --gpus 0, --scale_lr false

Model Zoo

Precomputed Databases

Source Dataset	Size [GB]	Link
OpenImages	18	https://ommer-lab.com/files/rdm/database/OpenImages/
ImageNet	1.2	https://ommer-lab.com/files/rdm/database/ImageNet/1281200x512-part_1.npz

Pretrained Autoencoding Models

For pretrained autoencoders see the latent diffusion repository.

Pretrained RDMs

Train-Datset	Train-Database	FID (Validation)	Precision	Recall	Link	Filesize [GB]
ImageNet	ImageNet	5.32	0.74	0.51	https://ommer-lab.com/files/rdm/models/rdm/imagenet_in-db/model.ckpt	6.2
ImageNet	OpenImages	12.28	0.69	0.55	https://ommer-lab.com/files/rdm/models/rdm/imagenet/model.ckpt	6.2
FFHQ	OpenImages	1.92*	0.93*	0.35*	https://ommer-lab.com/files/rdm/models/rdm/ffhq/model.ckpt	6.2

*: Evaluated using CLIP as feature extractor instead of Inception

Pretrained RARMs

Train-Datset	Train-Database	FID (Validation)	Precision	Recall	Link	Filesize [GB]
ImageNet-Dogs	OpenImages	45.27	0.64	0.55	https://ommer-lab.com/files/rdm/models/rarm/imagenet/dogs/model.ckpt	2.9
ImageNet-Mammals	OpenImages	49.92	0.56	0.58	https://ommer-lab.com/files/rdm/models/rarm/imagenet/mammals/model.ckpt	2.9
ImageNet-Animals	OpenImages	49.03	0.55	0.58	https://ommer-lab.com/files/rdm/models/rarm/imagenet/animals/model.ckpt	2.9

Get the models

All models listed above can jointly be downloaded and extracted via

bash scripts/download_models.sh

The models can then be found in models/{rdm,rarm}/<model_spec>.

Comments

Our codebase for the diffusion models builds heavily on OpenAI's ADM codebase and https://github.com/lucidrains/denoising-diffusion-pytorch. Thanks for open-sourcing!

BibTeX

@inproceedings{blattmann2022retrieval,
  title     = {Retrieval-Augmented Diffusion Models},
  author    = {Blattmann, Andreas and Rombach, Robin and Oktay, Kaan and M{\"u}ller, Jonas and Ommer, Bj{\"o}rn},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2022}
  doi       = {10.48550/ARXIV.2204.11824},
  url       = {https://arxiv.org/abs/2204.11824},
}

CompVis / retrieval-augmented-diffusion-models