Unstable Diffusion is an improved version of the Stable Diffusion neural network.

The authors believe that this will also improve the quality of generation.

Kickstarter has started a crowdfunding campaign called Unstable Diffusion. It’s an improved version of the Stable Diffusion neural network, uncensored.

Since version 2.0, Stable Diffusion has had more limitations: the neural network can no longer imitate the styles of some artists, generating explicit content with it has also become more difficult. The Unstable Diffusion team is going to fix it.

The project’s creators intend to train Stable Diffusion on their own datasets, removing the limitations. According to the authors of Unstable Diffusion, advanced training will not only allow the neural network to work with new requests, but also improve the quality of generation in general.

The Unstable Diffusion team explains this by using a rather unsuccessful dataset for basic Stable Diffusion training. After testing, the creators of Unstable Diffusion found that their dataset produced higher quality art.

Now Unstable Diffusion has a preview of the model. Here are some examples of the art that’s generated with it.

10.

Kickstarter is aiming to raise at least $25,000, and that goal is almost there. The money will go to training the model on a dataset of 75 million high-quality images:.

  • Approximately 25 million photos.
  • Approximately 25 million artworks from Artstation, DeviantArt and Behance.
  • Approximately 25 million cosplay and anime shots.

In addition, the funds will be used to complete the data set itself. It’s a time-consuming process, not only to select high-quality images, but to provide them with text descriptions — an important element of AI training.

The resulting model will be made publicly available, with the authors of Unstable Diffusion believing it will be available in January 2023. In the future, the model is planned to be adapted to new versions of Stable Diffusion.

If the campaign raises more than $25,000 (which it probably will), the additional money will go towards expanding the dataset, producing “mini-models” trained on a small number of very high-quality arts, and purchasing accelerators that community members will have access to.

The final of the crowdfunding campaign is scheduled for January 9, 2023.

Similar Posts