One scene.
Every checkpoint & strength.
Compare the untouched base with the Gold LoRA using matched prompts, seeds and sampling settings. Inspect style, composition, faces, hands and neutral controls.
Screening study, 16 prompts × 2 seeds. Validation loss is a subset measure; it does not establish visual quality. 125% is a stress test.
Validation curve & checkpoint provenance
The training validation subset uses 16 batches per rank at five fixed sigmas. The lowest observed mean is included as a candidate. Batch-level variance was not recorded, so no confidence interval is inferred from this curve.
Download full manifest · Frozen evaluation planHow to review & what comes next
Review prompt adherence, composition and counts, anatomy and hands, textures and artifacts, requested cinema style, and preservation of neutral controls. Check both seeds rather than selecting a single attractive result.
Shortlist up to three checkpoint/strength pairs. A separate blind review should use raters who have not browsed this labelled gallery, randomized sides and order, ties, repeated pairs and identical-image checks. This gallery is exploratory; do not present its best-looking cell as a proven quality gain.
Lock the candidate and settings before a larger evaluation with fresh prompts and seeds. The Gold test split remains unused during screening.