Try it out here! https://mobians.ai/
Join the discord for updates, share generated-images, just want to chat or if you want to contribute to helping perfect the model! https://discord.gg/KfAuP72uCd
USE:
The model uses Danbooru/E621 tags similar to anime models. I would not recommend just generating a character with no follow-up prompts as the results won't look very good. A good negative prompt also helps greatly.
Limitations:
Some characters generate less consistently. This problem will be consistently worked on as new versions progress.
Description
No more blurred edges!
Trained on ~4000 sonic images of various characters. fine-tuned character classifier and aesthetic grader were used to detect characters and filter images. https://huggingface.co/spaces/Metal079/Sonic_Character_tagger
This is a preview just to see how good it would look, final version will use ~15k images, be trained at 768x768 and have a more refined training set hopefully.
I'm proud to say it looks better than I expected!
Limitations: Due to the small-ish dataset, more niche characters are less consistent. This will hopefully be remedied in the full release.