RFD 1144: The image-to-OmniGen2 loop
Problem
Loop 2 is the smallest complete loop: OmniGen2 proposes an edited png, EditScore scores it, OmniGen2 repairs. One propose stage, one scorer, one repair arm, size M, and it runs second because the harness is the only thing it waits on.
Being smallest makes it the one that measures the two components every other loop also uses, so its numbers are the plan’s numbers.
Decision
Run loop 2 first among the four, and let it carry the measurements.
The hazard is a silent skip. The edit instruction comes from a text input rather than from the filename. A filename that matches no key is skipped without a word, which produces a run that reports success and edits nothing.
DETAILS.md holds what the first runs measured: what OmniGen2 costs at bf16 and at four bits, what EditScore costs, and what EditScore returns for an instruction the edit actually matched. fourloops-plan.usda and fourloops-etnf.usda both name that file as a source, so a quantity in either layer is checked against a measurement rather than against a recollection.