The Four-Card Product-Image Edit Brief

An AI-assisted product-image edit can sound harmless: keep the bottle, place it on a kitchen counter, soften the light. The risk is not only an awkward shadow. A polished candidate can quietly alter a label, cap seam, proportion, finish, or item count while looking ready for a catalogue. Current provider documentation describes masked or selected edits, whole-image edits, and reference-image workflows. These are capabilities, not a shared definition of what must remain true.

The brief is a proposed control for a synthetic example: one blue insulated bottle moved into a warm kitchen scene. Four cards fix product facts, limit editable surfaces, name continuity anchors, and make rejection explainable. Provider settings belong in the record, but they are not proof that product identity survived.

Card 1 — Freeze the product facts

Write the product boundary before the prompt. Record the SKU or variant, proportions, color, material, finish, logo and label wording, included pieces, and any buyer-relevant feature. Attach approved source photos, owner, version or date, and a hash where used.

Separate three states:

• Immutable: the bottle silhouette, cap construction, printed words, color family, and item count in the synthetic example.
• Editable: the background, countertop, ambient light, and a defined cast shadow, if those changes are allowed.
• Unknown: any surface hidden in the source or any detail not established by the product owner.

Do not let an AI-generated description become a product fact. Set tolerance per immutable field: logo lettering may be exact, a reflection may vary, and color may require owner judgment. Name the role that resolves an unknown.

Google’s current image-generation guidance suggests describing critical details such as a logo in detail when requesting an edit. That is prompt advice, not a guarantee of preservation. Treat it as an input aid, then compare the result against the locked source. Google’s Gemini image-generation and editing guide

Card 2 — Draw the editable surface

An editable surface is the named region where a change is permitted. Put the operation beside it: replace the background, extend the counter, alter ambient color, or add a soft shadow. List the prohibited region: product body, label, cap, hardware, and identity-bearing edges. Include a mask or selection overlay so another person can see the boundary.

OpenAI’s current image guide says its edits endpoint can use an image and mask to identify areas to replace. It also says masking is prompt-based, may not follow the exact shape, and requires matching format and size plus an alpha channel. The production conclusion is ours: a mask is guidance, not a pixel-level contract. OpenAI image generation and editing guide

Adobe’s docs distinguish partial-selection Generative Fill from whole-image Prompt to edit, and its Reference Image workflow supports object or whole-scene references plus swap or place intents. These controls do not prove an edit stayed in bounds. Adobe Photoshop, “Use reference images for consistent results” Adobe Photoshop, “Edit images with Prompt to edit”

If the change requires a whole-image operation, mark the extra risk instead of calling it background-only. Compare original and candidate with an overlay or blink review at full size. Inspect edges, fine print, seams, reflections, and contact points. An attractive scene can still fail.

Card 3 — Name the continuity anchors

Continuity anchors are the visual features that let a reviewer recognize the same product after the edit. Choose three to seven and mark each hard or flexible:

• silhouette, proportions, and orientation;
• camera perspective and scale in the frame;
• label placement, lettering, and logo geometry;
• material texture and characteristic highlights;
• contact plane, shadow direction, and shadow softness; and
• crop, color profile, and required negative space.

Hard anchors must survive within the product owner’s tolerance. Flexible anchors can adapt, but the allowed variation belongs in the brief. This is proposed review vocabulary, not a claim that a model measures continuity.

Reference images can improve the instructions a tool receives. Google documents multi-image editing and detailed descriptions for high-fidelity details; Adobe documents object or whole-scene references. Neither says a reference certifies unchanged pixels. A candidate may satisfy “warm kitchen” while changing perspective or putting the label on another plane. Read anchors against the original, not the prompt.

For the synthetic bottle, a passing candidate could change countertop and light while retaining front-facing angle, cap seam, label words, silhouette, and contact shadow. If the shadow changes the bottle’s apparent position, record a continuity failure even if the background looks natural.

Card 4 — Package rejection evidence

A rejection should be reproducible, not a matter of taste. Keep a small packet with:

1. the untouched original and its identifier;
2. the candidate, variant ID, and export dimensions;
3. the prompt or edit instruction and the mask or selection;
4. provider, model or app version, date, and relevant settings;
5. a full-size side-by-side plus annotated crops; and
6. the failed fact or anchor, disposition, and next permitted edit.

Useful labels include product fact changed, label unreadable, edge or halo, reflection contradicts material, shadow breaks continuity, unapproved surface changed, and source or permission unresolved. Preserve the rejected candidate as read-only evidence. Do not rename it “final.”

Content Credentials can add provenance and edit history. C2PA describes them as optional records and says provenance cannot establish that an image is true or factual. Treat credentials as context, not product verification. C2PA and Content Credentials explainer

Hold when the original is missing, the edit region cannot be recovered, a hard anchor fails, the run is untraceable, or the intended use needs a fact the image does not establish. Release only after independent readback at full resolution and intended display size. Describe the output narrowly: the approved product, in the approved scene, with listed surfaces changed.

A useful product-image brief does not say “make it better.” It says what is fixed, what may move, what must remain recognizable, and what evidence ends the attempt. That boundary gives AI room to explore without giving regeneration authority over the product.

Sources and limitations

• OpenAI, “Image generation” — checked September 10, 2026; supports image edits, reference images, masks, mask format and alpha requirements, and the limitation that GPT Image masking is prompt-based rather than exact. Limitation: provider documentation describes an API capability and does not prove that a particular candidate preserves product pixels.
• Adobe Photoshop, “Use reference images for consistent results” — checked September 10, 2026; supports object or whole-image reference use and swap or place intents in Generative Fill. Limitation: availability, model choices, and results depend on Photoshop version, account, selected model, and source image.
• Adobe Photoshop, “Edit images with Prompt to edit” — checked September 10, 2026; supports the distinction between whole-image Prompt to edit and partial-selection Generative Fill. Limitation: the interface distinction is not a guarantee that an edit stays within a selection.
• Google AI for Developers, “Gemini API image generation” — checked September 10, 2026; supports image-plus-text editing, multiple reference images, and guidance to describe critical details such as logos. Limitation: these are prompting and capability notes, not an accuracy or identity-preservation test.
• Coalition for Content Provenance and Authenticity, “C2PA and Content Credentials Explainer” — checked September 10, 2026; supports provenance and edit-history records, optional adoption, and the limitation that provenance cannot establish whether an asset is true or factual. Limitation: credentials can be incomplete or absent and require a compatible verifier.