So it turns out that I have to retrain images using the Flux.2 training model — this is in order to create loras that can be used with Flux.2 [dev].