Original content
I'm testing your model and it looks great.
Is the original dataset intact, or was part of it replaced when the anime and illustration-oriented content was included?
This model only adjusts the output of the MOSS-Image-SFT model (mainly by suppressing overly reactive layers) and does not involve any additional training. The dataset is identical to that of MOSS-Image-SFT.
MOSS-Image-SFT_clear is adjusted to remove noise from the final generated illustrations. It works well for anime-style illustrations, but for more realistic illustrations, the skin tones can sometimes become overly smooth.
One of the developers from the OpenMOSS team commented on Reddit that they are prioritizing the development of an anime version of MOSS-Image, so we may also be able to expect an official anime version of MOSS-Image to be released in the near future.