AI Product Image Quality Scorecard
A transparent 100-point rubric for deciding whether an AI-assisted product image is accurate enough to publish. Score the asset against its approved source, record the decision, and reject it when a protected fact changes.
- Version
- 1.0
- Published
- License
- CC BY 4.0
How to score an image
Review the generated image beside the approved source at full resolution. Award full points when a criterion passes, half points when the defect is visible but repairable, and zero when the criterion fails. Record evidence, not impressions.
- 01
Compare
Use the same source, brief, target crop, and final display size.
- 02
Score
Apply full, half, or zero points to every criterion.
- 03
Decide
Apply hard stops first, then use the total as a team threshold.
The 100-point rubric
The weights prioritize product truth over decorative polish. A beautiful background cannot compensate for altered packaging, impossible reflections, or an unsupported product claim.
Product fidelity
Does the generated asset still depict the approved product?
| Criterion | Points | Pass evidence |
|---|---|---|
| Silhouette and geometry | 8 | Shape, proportions, and major contours match the approved source. |
| Dimensions and scale | 6 | Relative dimensions remain credible and consistent with the source. |
| Components and controls | 6 | Caps, handles, seams, buttons, openings, and attachments remain present and correctly placed. |
| Label and logo placement | 5 | The location, size, and orientation of protected brand areas have not drifted. |
| Brand color and material | 5 | Approved colors and visible material characteristics remain recognizable. |
Physical realism
Do light, materials, perspective, and contact agree with one scene?
| Criterion | Points | Pass evidence |
|---|---|---|
| Contact shadow | 5 | The shadow touches the product and has a plausible softness and direction. |
| Light direction | 5 | Highlights, fill, and cast shadows agree with the stated source of light. |
| Reflections | 5 | Reflections correspond to the visible environment and do not invent contradictory objects. |
| Material response | 5 | Glass, metal, fabric, plastic, paper, and texture respond plausibly to light. |
| Perspective and surface contact | 5 | Camera perspective, horizon, product base, and supporting surface are coherent. |
Composition and delivery
Will the asset work in its intended placement without repair?
| Criterion | Points | Pass evidence |
|---|---|---|
| Crop and safe area | 5 | The requested aspect ratio protects the product and required platform safe areas. |
| Visual hierarchy | 5 | The product remains the primary subject and props do not compete with it. |
| Copy space | 5 | Reserved copy space is genuinely uncluttered at the final display size. |
| Edges and output quality | 5 | Edges, transparency, texture, and resolution survive full-size inspection. |
Factual integrity
Could the image mislead a customer about the product or brand?
| Criterion | Points | Pass evidence |
|---|---|---|
| Text and marks | 5 | No invented price, label text, certification, watermark, or altered trademark appears. |
| Claims and performance | 5 | The scene does not imply unsupported ingredients, results, scale, durability, or endorsement. |
| Props and context | 5 | Props support the brief without implying an unverified feature or included accessory. |
Rights and review record
Can the team explain where the asset came from and why it was approved?
| Criterion | Points | Pass evidence |
|---|---|---|
| Source permission | 4 | The team has permission to upload and reuse every source asset. |
| Human review | 3 | A named reviewer compared the output with the source at full resolution. |
| Decision record | 3 | Prompt, model, source version, output version, and approval decision are recorded. |
Worked scoring examples
These five hypothetical review situations explain how to apply individual criteria. They are teaching examples, not observed model results, customer cases, or a benchmark. Review every other criterion separately before calculating a total.
A readable label with one invented word
Text and marks: 0/5.
Reject regardless of the total. An invented label is a hard stop, even if its position, color, and background look correct. Compare every word with the approved source.
A detached contact shadow
Contact shadow: 0/5.
The shadow fails the contact requirement. Repair it and inspect surface contact and perspective separately; a high total does not remove the visible defect.
Too little space for the planned headline
Copy space: 2.5/5 if a small background edit can restore it.
Record the required safe area and repair the composition. If the product must be cut off or distorted to fit the headline, award zero instead.
A new handle on the product
Components and controls: 0/6.
Reject regardless of the total. The added handle suggests a feature the real product does not have. Return to the approved product geometry.
A polished image with no source permission record
Source permission: 0/4.
Reject until permission is established. Visual quality cannot substitute for evidence that the source may be used.
Examples added September 20, 2026. The version 1.0 rubric weights and downloadable files are unchanged.
Hard stops override the score
A high total does not make a misleading or unauthorized image publishable. Record the hard stop, reject the output, and return to the approved source or a controlled design workflow.
- A protected logo, label, certification, legal line, or price is invented or altered.
- Product geometry changes enough to suggest a feature that the real product does not have.
- The scene makes an unsupported factual, performance, ingredient, safety, or endorsement claim.
- The reviewer cannot establish permission to use the source image or included brand material.
- The delivered file hides a known defect through crop, blur, compression, or decorative overlays.
Suggested decision bands
These bands are editorial defaults, not measured industry benchmarks. Teams should publish their own threshold and keep it stable when comparing workflows.
Save the evidence behind your decision
Keep the source and output together with their filenames, model, prompt, and review date. Leave unverified fields marked unknown; do not turn an incomplete review into an approval. Use the CSV for criterion scores and this record for context.
Source file / version: [required] Output file / version: [required] Source permission evidence: [required] Model / version and generation date: [record or mark unknown] Prompt and settings: [record or mark unknown] Target placement and crop: [required] Reviewer and review date: [required] Completed scorecard file: [required] Hard stops: [not checked / none found / describe] Total: [enter only after all criteria are reviewed] / 100 Decision: [pending / reject / revise / candidate for approval] Evidence and required corrections: [describe] Re-review date and result: [if applicable]
Cite and adapt the rubric
The rubric text and downloadable files may be shared and adapted under the Creative Commons Attribution 4.0 International license. Attribute Create Image and link to this page. Product images, third-party marks, and source assets are not included in that license.