This model is trained collaboratively — it is a part of the NeurIPS 2021 demonstration "Training Transformers Together".

The latest model checkpoint will be uploaded to this repository every 6 hours until the training stops.

We train a model similar to OpenAI DALL-E — a Transformer model that generates images from text descriptions. Training happens collaboratively — volunteers from all over the Internet contribute to the training using hardware available to them. We use LAION-400M, the world's largest openly available image-text-pair dataset with 400 million samples. Our model is based on the dalle‑pytorch implementation by Phil Wang with a few tweaks to make it communication-efficient.

You can check our dashboard to see what is happening during the collaborative training (loss over time, number of active sessions over time, contribution of each participant, leaderboard, etc. ).

You can start from our Colab notebook for running inference.

This model is created only as a demonstration of the new distributed training methods. It should not be used for anything besides research purposes.

The authors have done only the most basic dataset filtering, so the model may be susceptible to biases in the training data and/or generate inappropriate content.

At the moment, the model's generative capabilities are limited due to the absence of extensive experiments with the architecture and incomplete training.