First attempt with a dataset I could never get to work correctly on Flux. Seems to work a bit better here.
v2.0 MOSS: retrained on MOSS-Image SFT and it finally works. It makes single-panel gag cartoons with clean, readable captions. Use the trigger "farside", describe one clear scene, and put the joke in a caption, e.g. ...At the bottom of the image is the caption "Still holding a grudge, I see."
Description
MOSS-Image version of the Farside Style LoRA, trained on MOSS-Image SFT. Works much better than the earlier attempts: single-panel gag cartoons with clean, readable captions.
How to prompt: start with the trigger and describe one clear scene, then put the joke in a caption, e.g. farside a black and white cartoon of a cow lying on a couch in a psychiatrist's office, while a bespectacled psychiatrist takes notes. At the bottom of the image is the caption "And how long have you felt that everyone is milking you?"
Without a caption the panels tend to be flat. Keep captions short and in plain words.
Settings: strength 1.0. Full quality: 50 steps, CFG 3, euler / sgm_uniform. The sample images use the MOSS SFT 4-step turbo LoRA (Kijai rank-72) stacked on top: 4 steps, CFG 1, euler / simple, about 9 s per image on a 2070 Super. Turbo is a bit cleaner and more generic than 50 steps and occasionally misspells unusual words, so render a few seeds.