Coming soon
No recorded evaluation run yet.
Proof is the post-training stage: it renders a seed/weight grid from a trained LoRA, scores every image (identity, CLIP alignment, aesthetics, preference, near-duplicates), gates the run on pass rate and diversity, and hands the borderline images to a keyboard-first human review. A real run will be recorded here once one has been made — this demo will not show invented results as if they were.