Teaching vision systems to adapt when the world changes

Perception systems deployed in the real world encounter shifting visual conditions—camera angles change, lighting varies, object appearances evolve. Standard continual adaptation methods try to align new data back to the original training domain, but this produces unreliable guidance. This work flips the strategy: instead, it pre-builds a bank of synthetic class examples, then dynamically adapts them to match incoming data's style at three levels (input, statistics, representation) while keeping their core meaning intact. The result is clean, trustworthy training signals that let models adapt stably through continuous shifts. Code released.