COLLABORATION BETWEEN LAION AND THE STABLE HORDE

Author: Konstantinos Thoukydidis, hlky

We are happy to announce that LAION will be assisted by the Stable Horde to provide aesthetic ratings for existing datasets and a completely new dataset of Stable Diffusion generations, which will also be rated by their community.

We wrote in the past about LAION-Aesthetics and how we filtered LAION-5b using an aesthetic predictor. The predictor, a simple neural net that uses CLIP ViT-L/14 embeddings as input. hlky has retrained the aesthetic predictor using ViT-H-14, and the results are promising.

hlky’s ViT-H aesthetic predictor was trained on the same datasets as the original - AVA: A Large-Scale Database for Aesthetic Visual Analysis, Simulacra Aesthetic Captions, and LAION-logos. These datasets are limited, totalling around 400k image-rating pairs. This is where the Stable Horde comes in.

The Stable Horde is a crowdsourced, distributed, free and open sourced service for generating Stable Diffusion images available to be integrated into everything. They have volunteered to direct their community to rate the existing LAION datasets for aesthetic rating of 1-10. As the Stable Horde can be used for free, to incentivize this, they are providing priority (AKA kudos) to anyone who is providing such ratings for their stable diffusion generations.

The integration has been made possible through an API hosted by Sygil.dev which can collect such ratings and be integrated directly into all Stable Horde clients. Already the first integrations have been released on clients like ArtBot.

Furthermore the Stable Horde has started gathering thousands of Stable Diffusion images and their parameters generated by their community on a voluntary basis, which will then be used to build a completely new AI dataset. These images will not only be incentivized to be rated aesthetically as well, but also to be rated pairwise.

Stable Horde supports many models in addition to base Stable Diffusion models. At the time of writing this, nearly 80 finetunes are available. There are currently no datasets of generated images from Stable Diffusion finetunes, this will be a world first!

The way this works is that whenever an image set (meaning a number of images out of the same prompt) is generated, a follow-up call can be done to submit ratings for all, or some of them. If the set includes more than one image, optionally the best of the set can be selected, which will serve as a pairwise rating. If no “best” image has been selected but they have been aesthetically rated, the best image will be automatically selected among those ratings.

Through this collaboration, LAION will be provided with a dataset of millions of images rated in such a manner, along with their prompts, which can be made freely accessible to train further models with improved quality.

What's next?

The Stable Horde plans to onboard more ratings into their dataset, such as tagging images for artifacts, watermarks, nudity and the like. The provided set will continue expanding so long as people volunteer their time and processing power.

Support this endeavour

If you have any questions or need support about the Stable Horde or the rating system, they have a very active discord server you can join. If you want to support this initiative, you can help by either rating images yourself, or by onboarding your own GPU as a horde worker which will help the community generate more images and thus increase the size of the open dataset they provide.