METHOD 0.1.0

Check the garment,
not just the image.

An attractive AI image can still change the product being sold. Compare it against the approved product photograph before using it in a listing, campaign or virtual try-on.

A useful comparison starts with the reference

1. Choose a sharp approved photograph with the garment fully visible.
2. Use the closest possible camera angle in the generated image. A flat-lay reference cannot establish how a hidden back seam looks on a model.
3. Inspect graphics, pockets, closures and edges first. Small changes can materially misrepresent a product.
4. Read the evidence and unverified attributes before accepting a score. Review lighting and pose whenever color or proportion is flagged.

What this beta measures

Twenty attributes span texture, color, details, construction and shape. Metadata, conservative plain-background foreground extraction, color samples, relative silhouette aspect and available OCR provide supporting measurements. Foreground extraction is not semantic segmentation. It is used as a garment comparison only when you confirm comparable isolated flat-lay photographs and both images pass basic background checks.

A vision model compares the images using an original structured rubric. Uncertain, structurally sensitive or conflicting cases receive a second model review when available. Material means visible appearance: fiber composition cannot be verified from a photograph.

How the result is formed

The score is an equal-weight mean of visible attribute scores, rounded to a whole number. Not applicable and not visible attributes are excluded. Coverage is visible attributes divided by applicable attributes. Confidence is the model’s self-assessment, not a calibrated probability.

• PASS: score at least 90, coverage at least 75%, confidence at least 80%, comparable views, no visible attribute below PASS and no unresolved computer-vision conflict.
• FAIL: at least one visible attribute scores below 70 with confidence at least 80%.
• REVIEW: all other cases, including missing vision service. An unresolved garment identity receives no overall score.
• NOT VERIFIABLE: the attribute is hidden, unreadable or inapplicable; the evidence states why.

Computer-vision disagreements trigger escalation and block an automatic PASS. They do not independently prove a garment defect. These initial thresholds are product heuristics, not validated certification criteria. No benchmark or accuracy claim is made.

Research context

The use of explicit apparel attributes is informed by Park, Choi and Choi (2026), Fashion and Textiles. Their work motivates granular visual review and highlights limitations in structural judgments. This beta uses its own prompts, decision rules and computer-vision checks. It is not affiliated with the authors and has not reproduced their validation study.

Privacy and retention

SKUExact does not train on uploads. Images are handled in memory and, if the upload parser or OCR requires a temporary file, in the service’s temporary directory. Request cleanup closes uploads; deployment uses memory-backed temporary storage. No image, filename, OCR text or visual evidence is saved in the analytics database or normal application logs. Browser previews remain until you clear them or leave the page.

When vision is enabled, you must consent before your images are sent to OpenAI. Requests use store: false. OpenAI API data is not used for training by default, but abuse-monitoring retention can still apply, commonly up to 30 days; this app does not promise zero provider retention. See the provider’s data controls. Diagnostics-only mode does not send images to a model provider.

We keep minimal result counts, numerical scores, elapsed time, model token use and generic error codes for 30 days, pruned hourly and at startup. A daily salted IP hash enforces free limits; raw IP addresses are not stored in the app database. Downloaded reports are yours to retain and can include recognized text. Application access logging is disabled in the dedicated proxy and app service.

Free beta, with future plans

The current beta has a limited daily allowance and no payment requirement. Pro and Studio are planned for larger allowances, batches and projects. Checkout, affiliate links and public benchmarks are inactive.