AI product image consistency

Treat AI product image consistency as a deliverable, not a definition. The job receives one master SKU photo and three target angles and owes a matched three-angle SKU set in 1:1 and 4:3. Write the rejection reason first, then turn it into the first hard constraint.

Do not start until the source photo has a visible product code, a note on material and color, the approved claim, and a named target module. Check packaging damage, sticker state, lighting direction, and background rights before generation. Do not invent certifications, competitor claims, price windows, or a model using the product in an unsafe way. If the seller cannot prove it, the image does not carry it.

Open Asset Library first, attach the master SKU as the locked reference, then build the module or shot around that reference. Keep offer copy, platform text, and CTA outside the visual prompt. In imgmov, the approved image becomes the source of truth and generation supplies only the requested scene, angle, or motion.

Keep platform requirements outside the model prompt. The prompt can describe a scene; a named checklist records the crop, claim, consent, and console check that make it deliverable.

Lock product silhouette, label spelling, logo position, material, primary color, accessory count, contact shadow, and the platform safe area. Scene, props, and motion may change; these fields may not.

Put the locked fields at the top of the prompt, not at the end: The practical move is to lock the exact product shape, label, color, and material and change only three angles from one locked reference. Save that prompt with the reference, ratio, and model so the next run starts from a decision instead of a guess.

1) Set the rule: I opened the one clean SKU photo and one selling point, locked the exact product shape, label, color, and material, and wrote down what could not change. Leave one checkable artifact from this step; do not start the next until it exists. 2) Run one minimum version: I made one small a matched three-angle SKU set first. The side view changed proportions, so I locked the front view as the master reference. Leave one checkable artifact from this step; do not start the next until it exists. 3) Build the focused variant: I reused the same reference and changed only three angles from one locked reference. Leave one checkable artifact from this step; do not start the next until it exists. 4) Review and export: I checked text, edges, shadows, color, and final crop at full size, then exported the final 1:1 and 4:3. Leave one checkable artifact from this step; do not start the next until it exists.

Change one named variable per rerun: prompt, reference, camera, duration, model, ratio, or export crop. If two variables change together, a better output cannot be reused because nobody knows which fix worked.

Review at 100%. Read small text, count buttons and parts, compare color against the source, inspect the four corners, and overlay the platform crop. A thumbnail is only a concept check, not a marketplace check.

Keep source, approved wording, rejected version, correction, credit cost, and final crop next to the asset. The record exists so the next reviewer can reproduce the decision without a chat thread.

Name the failure first: label redraw, color drift, shape drift, missing accessory, wrong contact shadow, or cropped CTA. Stop prompting for more letters, reduce the variable count, return to the reference, and only then upgrade the model.

Image retries have a ladder. Start with Agnes Image 2.5 for the cheap proof. Use SDXL Inpainting at 3 credits when only a masked region is wrong. Use Seedream 5.0 Lite or Nano-Banana 2 at 5 credits for a more complex rebuild. GPT Image 2 starts at 2 credits for Standard 1K and rises to 80 by quality and resolution. Name the failure before changing rungs.

Deliver by module and ratio, not one stretched master. Name files with SKU, module, crop, version, and approval state. Keep the rejected render and the correction line in the same folder.

The folder is boring on purpose: source, approved reference, generation settings, caption file, platform cuts, QA screenshots, and one correction line. A reusable asset is boring in the right way.

Editing is not detached from production. Object detection, masks, reference images, version history, Canvas dependencies, and the final generation set stay together. A standalone generator can invent a scene; it will not preserve the source asset, the rejected versions, and the reason the approved crop won.

Do not invent certifications, competitor claims, price windows, or a model using the product in an unsafe way. If the seller cannot prove it, the image does not carry it. If the request only asks what ai product image consistency means, a search page is faster. This page is useful when someone must deliver a matched three-angle sku set under real constraints.