The source photos were unused in the available prior project records. The catalog products were already used in development. This small study tests new photos of known products; it does not establish performance on shoppers or new products.
4 / 4plain-top alternatives within ΔE 10.0; 0 missing
2patterned reference tops excluded from single-colour scoring
2source photos, with three correlated renders each
0independent human rubric reviews; requested change assessed in 0 / 4 comparisons
12 unitsreconciled against task history; 958 remaining at the post-run check
Colour closeness is not overall fidelity. An unblinded AI observation flags pushed-up sleeves in one sage-top result. The catalog says long sleeves; this image does not verify the requested coverage. Independent product-detail and requested-change ratings remain unassessed.
Executed 2026-10-29T04:09:35.633530+00:00. No retries, corrections or selected rerenders. All six first outputs are below. No combined accuracy score is calculated.
Elapsed time includes upload and polling. Two units were charged per task. Development measurements retain their own denominators.
Reproduce the measurements
The design, measurement code and thresholds were committed before generation. Reporting glue was added afterwards. This command checks the frozen hashes and recomputes measurements from the exact public normalized inputs and stored output JPEGs, without an API key, database data or private images.
The immutable registration still says “not run”: it records the pre-run plan, not current execution status. Results are recorded separately here. Private task identifiers are hashed in the public record.
Exact normalized input photoCatalog productUncorrected first output
Colour: close, ΔE 5.5. Colour check: close to the product photo.
Outside-region pixel change: 0.0%. Fixed approximate boxes and threshold 40; a zero does not establish identity, pose or background preservation.
AI observation, not blinded. The sleeves appear pushed up on this render, exposing forearms despite the long-sleeve product. A close colour measurement does not verify requested sleeve coverage. This is an unblinded AI observation, not a human rubric rating. Reviewer: Codex (AI assistant).
Independent human review pending. Product detail, requested difference and misleading-result judgments are unassessed.
Exact normalized input photoCatalog productUncorrected first output
Colour: close, ΔE 2.5. Colour check: close to the product photo.
Outside-region pixel change: 0.0%. Fixed approximate boxes and threshold 40; a zero does not establish identity, pose or background preservation.
Independent human review pending. Product detail, requested difference and misleading-result judgments are unassessed.
Limits
Two synthetic source photos and known catalog products; this does not establish shopper, new-product or demographic performance.
No automatic retries were made; every frozen case and its first attempt is included.
No independent human rubric reviews have been completed. The one sleeve note is an unblinded AI observation.
The original feasibility runner includes an old ten-pair pilot gate. That gate is not used for this six-case study; the frozen held-out protocol governs its denominators and conclusions.
Reported outputs are the app-stored JPEGs. Source inputs match the pre-registered normalized hashes; output bytes are copied without further resizing or editing.
Client elapsed time includes upload and polling waits; it is not model-compute time.
Colour closeness and successful task completion are separate observations, not overall fidelity or physical-fit scores.
Outside-region changes use the original fixed boxes and threshold. A zero value is not evidence of identity, pose or background preservation.
Human product-detail, requested-change and misleading-result judgments remain unassessed.