Flux 2 Klein

Black Forest Labs released Flux 2 Klein as their distilled, latency-first image model, and this skill wraps the 9B and 4B variants on RunComfy so Claude can hit sub-second generation when you're iterating. The built-in prompting guidance is the real value here: it documents the subject-first declarative grammar the model was trained on, step-count strategy by phase (4 steps for concepting, 25 for polish), and when the 4-step distillation breaks down. The skill knows to route you elsewhere when you need 4K resolution or heavy text rendering. If you're doing live art direction or rapid product visualization where feedback speed matters more than ceiling quality, the 4B variant at 4 steps is legitimately faster than waiting for batch-style models to return.