Home

Awesome

ImageBART

NeurIPS 2021

teaser <br/> Patrick Esser*, Robin Rombach*, Andreas Blattmann*, Björn Ommer<br/> * equal contribution

arXiv | BibTeX | Poster

Requirements

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

conda env create -f environment.yaml
conda activate imagebart

Get the Models

We provide pretrained weights and hyperparameters for models trained on the following datasets:

Download the respective files and extract their contents to a directory ./models/.

Moreover, we provide all the required VQGANs as a .zip at https://ommer-lab.com/files/vqgan.zip, which contents have to be extracted to ./vqgan/.

Get the Data

Running the training configs or the inpainting script requires a dataset available locally. For ImageNet and FFHQ, see this repo's parent directory taming-transformers. The LSUN datasets can be conveniently downloaded via the script available here. We performed a custom split into training and validation images, and provide the corresponding filenames at https://ommer-lab.com/files/lsun.zip. After downloading, extract them to ./data/lsun. The beds/cats/churches subsets should also be placed/symlinked at ./data/lsun/bedrooms/./data/lsun/cats/./data/lsun/churches, respectively.

Inference

Unconditional Sampling

We provide a script for sampling from unconditional models trained on the LSUN-{bedrooms,bedrooms,cats}- and FFHQ-datasets.

FFHQ

On the FFHQ dataset, we provide two distinct pretrained models, one with a chain of length 4 and a geometric noise schedule as proposed by Sohl-Dickstein et al. [1] , and another one with a chain of length 2 and a custom schedule. These models can be started with

CUDA_VISIBLE_DEVICES=<gpu_id> streamlit run scripts/sample_imagebart.py configs/sampling/ffhq/<config>

LSUN

For the models trained on the LSUN-datasets, use

CUDA_VISIBLE_DEVICES=<gpu_id> streamlit run scripts/sample_imagebart.py configs/sampling/lsun/<config>

Class Conditional Sampling on ImageNet

To sample from class-conditional ImageNet models, use

CUDA_VISIBLE_DEVICES=<gpu_id> streamlit run scripts/sample_imagebart.py configs/sampling/imagenet/<config>

Image Editing with Unconditional Models

We also provide a script for image editing with our unconditional models. For our FFHQ-model with geometric schedule this can be started with

CUDA_VISIBLE_DEVICES=<gpu_id> streamlit run scripts/inpaint_imagebart.py configs/sampling/ffhq/ffhq_4scales_geometric.yaml

resulting in samples similar to the following. teaser

Training

In general, there are two options for training the autoregressive transition probabilities of the reverse Markov chain: (i) train them jointly, taking into account a weighting of the individual scale contributions, or (ii) train them independently, which means that each training process optimizes a single transition and the scales must be stacked after training. We conduct most of our experiments using the latter option, but provide configurations for both cases.

Training Scales Independently

For training scales independently, each transition requires a seperate optimization process, which can started via

CUDA_VISIBLE_DEVICES=<gpu_id> python main.py --base configs/<data>/<config>.yaml -t --gpus 0, 

We provide training configs for a four scale training of FFHQ using a geometric schedule, a four scale geometric training on ImageNet and various three-scale experiments on LSUN. See also the overview of our pretrained models.

Training Scales Jointly

For completeness, we also provide a config to run a joint training with 4 scales on FFHQ. Training can be started by running

CUDA_VISIBLE_DEVICES=<gpu_id> python main.py --base configs/ffhq/ffhq_4_scales_joint-training.yaml -t --gpus 0, 

Shout-Outs

Many thanks to all who make their work and implementations publicly available. For this work, these were in particular:

teaser

References

[1] Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N. & Ganguli, S.. (2015). Deep Unsupervised Learning using Nonequilibrium Thermodynamics. <i>Proceedings of the 32nd International Conference on Machine Learning

Bibtex

@article{DBLP:journals/corr/abs-2108-08827,
  author    = {Patrick Esser and
               Robin Rombach and
               Andreas Blattmann and
               Bj{\"{o}}rn Ommer},
  title     = {ImageBART: Bidirectional Context with Multinomial Diffusion for Autoregressive
               Image Synthesis},
  journal   = {CoRR},
  volume    = {abs/2108.08827},
  year      = {2021}
}