generation prompt revision diff: Review the change as a field-level hypothesis

A generation prompt revision diff should explain what changed in one proposed request and why that change is being tested. It is a review tool for the next experiment, distinct from the broader task of storing prompt versions. A useful diff makes accidental changes visible before submission.

Compare the entire approved request

Start with the previous approved request and the proposed request. Compare model, endpoint, prompt, ratio, duration or image count, resolution where applicable, seed, reference selection, negative prompt, and extra.

A prompt-only review can miss a changed model or image count. Those changes can affect both the experiment and its credit reservation. Put them beside the wording difference so the reviewer sees the full proposed work.

If your brief is organized into local fields such as subject, camera, lighting, and composition, compare those fields before compiling the final prompt string. They are application fields; the gateway still receives prompt as text.

Name one hypothesis

For a fictional product image, the old composition clause might say product centered on a textured surface. The revision might say product on the right with a plain upper-left area for a later headline. The hypothesis is that the planned overlay region becomes more usable.

Record exactly which clause was removed and which was added. Keep lighting, subject identity, model, ratio, and sample plan unchanged if the experiment is intended to isolate composition.

If multiple changes are necessary, label the revision as a combined change. It can still be useful, but a later improvement cannot be attributed confidently to one clause.

Review semantic conflicts

Read the compiled string after examining the field diff. A new camera instruction can conflict with an older sentence that remains elsewhere. A replacement of soft light with strong directional light can leave incompatible shadow instructions behind.

Check the final serialized request too. Aliased standard fields can contain contradictory values, and extra can override upstream mappings. A clean prose diff does not guarantee an unambiguous HTTP payload.

Confirm that seed zero remains present when intended and that the selected first reference has not changed through upload ordering. Those are request differences even when the creative text is identical.

Define the observation before submission

State what a reviewer will inspect and what would count as an improvement or failure. For the overlay example, use the actual headline in the intended crop rather than judging empty space by impression alone.

Leave results blank until real artifacts are collected and reviewed. This proposed procedure does not demonstrate that any wording improves model quality.

Approve the diff when its changed fields match the stated hypothesis and its cost exposure is understood. Continue with prompt versioning, negative-space prompts, and pricing snapshots before authorizing the next sample.

Continue with quality and testing

• generation model regression checklist: Maintain a release scorecard with real evidence
• AI image color drift test: Control viewing conditions before judging color
• image generation evaluation rubric: Give human scores observable anchors
• Browse all quality and testing guides