An AI-generated scene can look polished while quietly changing the product. A bag gains another pocket, a bottle label becomes unreadable or an accessory appears in the offer even though it is not included. The review needs to check what the image communicates, not only whether it looks good.
Use this process after the HeliosGen and Kie.ai setup guide. Download the AI asset approval sheet. The commuter-backpack example below is a proposed exercise; we have not generated or tested campaign results for it.
Assemble a reference that can be checked
Start with authorized product photographs and the actual product details. Include a front view, a side view and close-ups of any feature the ad will show. Record the exact product version, available colors and included accessories so a reviewer does not compare an old reference with a new SKU.
Write three lists: features that must remain accurate, scene elements that may change and claims that require separate evidence. For a backpack, the pocket placement and strap construction belong in the first list. The desk surface and background lighting may belong in the second. A statement about water resistance belongs in the third unless you have appropriate support for it.
If there is no reliable reference for an important detail, resolve that gap before asking a model to fill it in. A plausible detail is still a made-up detail.
Keep the first generation small
Kie's Nano Banana 2 documentation exposes image inputs, aspect ratio, resolution and output format. Those controls describe the request; they do not guarantee product fidelity. Record the settings supported by the specific provider and tool version you are using.
In HeliosGen, first make one output using the connected provider and inspect it before duplicating branches. Keep the input reference, prompt, model and task or export identifier together. The existing setup guide provides the version-specific connection and workflow instructions, including the distinction between saving a key and completing a successful generation.
An original starting brief could be:
Create a neutral desk scene around the supplied backpack reference. Preserve the visible product shape, pocket layout, straps and color. Do not add accessories, text, logos or a person. Leave clear space above the product for separately designed copy. Treat the reference as the product specification.
This is a request to the model, not an assurance that it will comply. If the result changes the product, reject or correct it before using it in an ad.
Review the product at useful magnifications
Inspect the full composition, then enlarge the product. Compare both against the reference. Check seams, closures, handle geometry, repeated details, logo placement and any visible label. For products with precise packaging text, a verified photograph or approved composite may be more suitable than regenerating the packaging itself.
Use a simple approval rule: a product-critical mismatch blocks release. Do not let a high overall aesthetic score compensate for a changed feature. Put the issue in the sheet with an owner and the exact file version that needs correction.
| Check | Example failure | Required decision |
|---|---|---|
| Product structure | A second front pocket appears | Reject or correct the asset |
| Included items | Headphones appear beside the bag as if bundled | Remove them or clarify the offer |
| Visible text | The logo is distorted | Use verified artwork or another source |
| Scale and use | The bag is implausibly small on a desk | Correct the scene and review again |
Check the crop and the claim together
Preview every intended placement using its actual current specifications. A source image with empty space can still crop badly when a placement changes the visible area. Confirm that the product, essential text and call to action remain readable on a phone.
Add final advertising copy after visual approval and review it as part of the assembled ad. A neutral scene accompanied by “fully waterproof” still makes a claim, even if the generation prompt never mentioned water. Verify the landing page offers the same product and includes the relevant conditions.
Keep provenance and disclosure checks in the release record. Meta's AI advertising transparency update describes expanding information about AI-created or edited ads. Review the current platform controls for the actual asset rather than assuming a file export handles disclosure automatically.
Hand off a traceable approved version
Save the approved export under a stable asset ID, with the reference, generation settings, edits, reviewer and date. Keep rejected versions separate so they cannot be selected accidentally during assembly.
Then attach the asset to a defined creative concept and test budget. The generation process produces material for a campaign; the approved brief and measured results determine whether that material is useful. Take that evidence into a supported AdRails review.
Sources checked October 14, 2026. This quality checklist does not run a paid generation or publish an ad.