EveryDream SD1.5 768

This is a fine tuned version of the SD1.5 checkpoint and MSE VAE initially trained by RunwayML. This is further fine tuned to produce reliable images at a nominal resolution of 768×768 or other aspect ratios of the same total pixel count.

Trained using EveryDream2 (https://github.com/victorchall/EveryDream2trainer) with approximately 30,000 hand picked images for 40 total epochs starting at 512 resolution, then 640, and finally 768. Training data consists of a wide variety of imagery from fine art to video games with no particular focus. No NSFW images were used to train this model.

This model is suitable for continuation training for those who want to start at higher resolutions with the SD1.x architecture. You can use or resume training simply using “panopstor/SD15-768” with Diffusers enabled applications.