Product-fidelity review sheet
Creator Lane | Reviewed September 20, 2026
https://creatorlanehq.com/blog/ai-product-images-accurate-labels

PRODUCT-FIDELITY WORKSHEET

BLANK REFERENCE BRIEF
Product and exact variant:
Approved source files:
Permission / owner:
Exact label copy:
Shape and dimensions reference:
Visible parts:
Material / finish:
Colour reference:
Allowed scene changes:
Details that must not change:
Unsupported features or claims to exclude:

EDIT PROMPT
Use [source] as the authority for [product]. Change only [surroundings or limited area]. Build [scene description] with [light direction]. Leave space at [location] for editable text. Preserve the visible product as closely as possible. Do not invent [specific prohibited details]. I will check the result against the source before publication.

REVIEW EACH CANDIDATE
Candidate filename:
Actual model / interface / date:
Reference version:
Silhouette:
Parts:
Label:
Material:
Colour:
Scale:
Interaction:
Implied claims:
Repair needed:
Use original cutout instead?:
Reject reason:
Reviewer:

FILLED PLANNING EXAMPLE — fictional, not a tested result
Item: matte green travel bottle.
Locked facts: upright body, dark screw cap, exact approved label.
Permitted change: off-white tabletop setting.
Text area: right side, added in editor.
Primary risk: changed label or cap.
Fallback: original bottle cutout over empty background.
Candidate / test result: none generated for this example.

REPEATABLE BATCH ACCEPTANCE PROTOCOL
Batch ID / purpose / final destination:
Source-reference version / approved product variant:
Actual model identifier / app or interface / date:
Prompt version / available output settings:
Maximum attempts / spending limit chosen for this batch:
Mandatory pass conditions agreed before generation:
Person authorized to approve product facts:
Permission and publication reviewer:

FOR EVERY CANDIDATE — duplicate this record
Unique filename / candidate ID / revision ID:
Source / prompt version:
Actual generation charge, if known:
Review and repair minutes:
Silhouette: pass / fail / not verified
Parts and closure: pass / fail / not verified
Exact label characters and symbols: pass / fail / not verified
Material / finish / colour variant: pass / fail / not verified
Scale / reflection / shadow / interaction: pass / fail / not verified
Unsupported features or implied claims: pass / fail / not verified
Final crop / size / editable text: pass / fail / not verified
Source and added-asset permissions: pass / fail / not verified
Product reviewer / date / evidence:
Publication reviewer / date / evidence:
Decision: accept / reject / hold
Specific reason and next action:
Final deliverable ID if accepted:
A failed or unverified mandatory check is not accepted.
Every repair gets a new version and the full checklist again.
Count distinct final deliverables, not multiple versions of one image.

FICTIONAL DECISION TABLE — no images generated or tested
C01: required product details pass; publication checks pass; ACCEPT.
C02: required product details pass; publication checks pass; ACCEPT.
C03: label character changed; REJECT.
C04: cap shape changed; REJECT.
C05: wrong colour variant; REJECT.
C06: visible product details pass, background permission unverified; HOLD, not accepted.
These are invented decisions, not measured model results or an expected success rate.
Later revisions or cleared holds require a dated decision before changing the accepted count.

COST PER ACCEPTED OUTPUT — enter actual inputs for your batch
Total generated candidates, including rejects:
Accepted distinct final outputs:
Generation spend and currency:
Other included cash costs / definition:
Total review and repair minutes, including rejected candidates:
Chosen hourly value of time:
Time value = minutes / 60 x hourly value
Cash cost per accepted output = included cash costs / accepted outputs
Combined cost per accepted output = (included cash costs + time value) / accepted outputs
Time value is not automatically a cash expense. Avoid double counting paid review.
If accepted outputs = 0, report UNDEFINED / no accepted outputs, plus actual spend and time.
Do not replace zero accepted outputs with one or report a zero unit cost.

FICTIONAL COST EXAMPLE — not provider pricing or a model test
Candidates: 6
Generation spend: INR 120
Accepted final outputs: 2
Generation spend per candidate: INR 120 / 6 = INR 20
Generation spend per accepted output: INR 120 / 2 = INR 60
Review: 30 minutes at a chosen INR 600/hour = INR 300 time value
Combined batch cost: INR 120 + INR 300 = INR 420
Combined cost per accepted output: INR 420 / 2 = INR 210
Additional cash costs assumed zero in this example, not necessarily in your project.
No accepted output? Unit cost undefined; preserve the spend and time totals.

NEXT BATCH / FALLBACK
Recurring rejection reason:
Change one documented method or reference:
Use original product cutout / photograph instead?:
Repeat all checks after model, product, prompt or workflow changes:
Do not carry a previous batch's approval onto new outputs.
Official capability reference reviewed September 20, 2026:
https://ai.google.dev/gemini-api/docs/image-generation

THREE SIGN-OFFS
Product accuracy: name / date / notes
Creative quality: name / date / notes
Publication readiness: name / date / notes
Final export:
Source and project location:


You may adapt this original worksheet for personal and client work. Third-party materials retain their own licences.