RFD 1145: The stylized-to-OmniGen2 loop
Problem
Loop 3 is loop 2 with a style transfer in front of it: CycleGAN proposes a restyled image, OmniGen2 edits it, EditScore scores the result and OmniGen2 repairs. The artifact is an edited png and the size is M, the same as loop 2.
What makes it a separate loop rather than a variant is where its input comes from, and that is a licensing question before it is a technical one.
Decision
Run it, and treat everything it touches as evaluation only.
The hazard is the holdout. The COCO-OOD stylized sets are val2017 restyled, so they are held out twice over: derived from the blinded holdout, and generated. CLAUDE.md states the rule this loop is most likely to break — anything derived from val2017 inherits its status, and a set generated from a held-out photo carries that photo’s content wherever it goes.
So loop 3 produces no training data, whatever it produces. A stylized corpus that reaches a training run has taken the holdout with it, and nothing downstream would report that it had.