reference image fidelity test: Inspect identity and geometry against an approved source

A reference image fidelity test asks whether an output preserves the specific details required by your brief. It does not assume that a reference-assisted generation is a faithful reproduction. Define the details that matter, then compare them with an approved source under a consistent review method.

Establish the actual reference

Retain the source asset, its content hash, approval status, and the exact URL position sent in the request. The current adapters normally map only the first imageInputs entry to image_url. Several uploaded files do not establish that every view influenced generation.

Confirm the serialized request before evaluating fidelity. If a side-view reference remained in the interface but the front view was the first submitted image, the review should name the front view as the generation reference.

A valid URL also does not prove that the intended bytes remained accessible upstream. Keep source provenance and output review as separate evidence.

Identify distinguishing features

Make a short checklist from the subject: outline, major proportions, number and placement of components, material boundaries, markings, and details that distinguish the item from similar objects. Add tolerances or categorical rules appropriate to the use.

For a fictional travel mug, a reviewer might check the handle shape, lid mechanism, base width, and the position of a colored band. A pleasant generic mug is not sufficient if those features define the actual product.

Do not require details that are absent or invisible in the source without another approved reference. Mark them as unknown rather than inventing a target from memory.

Compare in the right coordinate context

Inspect the source and output side by side. Account for viewpoint and framing before interpreting a size difference as changed geometry. Where practical, compare normalized proportions between identifiable landmarks rather than raw pixel distances from differently cropped images.

Then inspect critical regions at a useful zoom and at the final delivery size. An output may resemble the source overall while altering one small but important component. Record the location of each mismatch so another reviewer can find it.

Do not claim a universal fidelity percentage from this checklist. Unless a measurement method and tolerance are defined and actually applied, the result is a documented human judgment for the named task.

Decide the use boundary

An asset can be acceptable for an early visual concept and unacceptable for a factual product listing. Record that intended-use decision explicitly. When exact geometry or branding is required, approved photography or deterministic composition may be necessary.

Accept the review record when it identifies the submitted reference, visible evidence, required features, uncertain regions, and final use decision. Continue with multiple reference inputs, human review rubrics, and output storage to preserve the complete comparison.

Continue with quality and testing

• generation model regression checklist: Maintain a release scorecard with real evidence
• generation prompt revision diff: Review the change as a field-level hypothesis
• AI image color drift test: Control viewing conditions before judging color
• Browse all quality and testing guides