Metal079/SonicDiffusion

UPDATE: V2 is released as is much better, would recommend using instead https://huggingface.co/Metal079/SonicDiffusionV2

3 Dreambooth models based on AnythingV3 for the base model and training images from Evan Stanley's twitter.

evan5400 was trained on ~30 images for 5400 steps, use keyword 'sonic person' when prompting mobianstrimmed6000 was trained on 100 images for 6000 steps, use keyword 'mobian person' when prompting mobianstrimmed12000 was trained on 100 images for 12000 steps, use keyword 'mobian person' when prompting

Current testing shows that mobianstrimmed6000 and evan5400 produce the best quality images, I would recommend starting with one of those two and compare results.

Apologies for using different keywords between models, I wanted to fully switch over to 'mobian person' but unfortunetaly im not fully conviced even though I used more than double the training images, that it is better so I included both models. 6000 seemed like the sweet spot as I have noticed mobianstrimmed12000 doesnt give as consistantly good images as the other two.