PieMOSS is a fine tune of MOSS-Image trained on 15k datapoints.

Version 0.1: Initial version

Version 0.2: More high quality datasets added (23,218!)

Mid Nov - Version 0.3: Training specific to background crowds and background text/details, NSFW datasets, char Lora reinforcement

Goal: Avoid photos with the AI "polished" look. This is subjective and data was mostly pruned by me.

Implementation: Basically I babied it. Standard training with regression checkpoints. Most regression issues were around losing text (signs, tshirts, etc) so had to re-train and adjust settings every few thousand steps to accommodate.

Known issues: MOSS, in general, has trouble with background details. Particularly text, faces, and things that aren't directly prompted for. PieMOSS has the same issues so far.

NSFW Capability/Censorship: This model was NOT trained on NSFW concepts nor data. I fear the datasets/collections are poisoned with AI generated data for those concepts and it will be too much work for me to prepare the data correctly right now. Thus, this model should have similar capability to the SFT MOSS model.

Settings

Accelerator: Use 8 step dmd2 lodestone at 10-12 steps; euler simple, er_sde simple, euler beta

Baseline: Use 30-50 steps

Version goals:

PRERELEASE 0.2 - Better realism and additional datasets

PRERELEASE 0.3 - Focus completely on improving background details such as faces/text/wall patterns

PRERELEASE 0.4 - Better realism and additional datasets

PRERELEASE 0.5 - Skin Detail finetuning

FULL RELEASE 1.0 - Additional steps based on either 0.4 or 0.5, whichever is better