Make AI Product Visuals Without Losing the Label, Shape or Colour

To make AI product images with accurate labels, start with approved photographs, limit the edit, and check every candidate against the same product truth sheet. Accept only outputs that pass factual and publication checks, then calculate cost per accepted image—including review time—rather than judging the cheapest generation.

The most reliable starting point is an owned photograph of the actual item. Decide what can change around it and what must stay true. This guide provides original editing prompts and a review worksheet. It does not present generated examples or claim that a model passed a product-fidelity test.

Choose the smallest change that solves the problem

There are three different jobs that people call “AI product photography.” Pick yours before prompting.

A new background: the actual product photo is already good. Cut out the product and place it into a new setting. This keeps the most important pixels under your control.

A limited photo edit: the angle is useful, but the surrounding table, lighting or empty space needs work. Ask for a constrained edit and inspect whether the product changed as a side effect.

A newly generated scene: the camera angle or interaction does not exist in your references. This gives the model more freedom and requires much stricter review. It may be suitable for a clearly illustrative concept, but not for proving a physical feature you have never photographed.

Google documents image editing with supplied references. That is an available capability, not a guarantee that every label, seam or dimension will survive an edit. Google image-editing documentation

Build a product truth sheet

Imagine a fictional product used only for this worksheet: a matte green travel bottle with a dark screw cap and a small printed label. Do not treat this description as evidence about a real item.

For your actual product, collect a front photo, a side view and a close-up of important details. Use consistent lighting where possible. Keep the original files. Screenshots taken from a marketplace are a poor substitute for a reference you are entitled to use and can inspect closely.

PRODUCT TRUTH SHEET — fill from approved facts
Product / variant:
Source photographs:
Who owns or approved the photographs:
Front label, exact wording:
Cap / closure:
Number and position of visible parts:
Material and finish:
Colour reference:
Features that must not appear:
Claims allowed in this campaign:
Person who can approve product accuracy:

Separate visible facts from marketing language. “Two zip pockets” is something a photograph may demonstrate. “Lasts forever” is a claim an attractive image cannot establish.

Write the edit boundary in ordinary language

Tell the model what the reference is for, what to change and what to leave alone. Use a clean image with one obvious subject. If you supply several variants, label them clearly in your instructions so the model does not combine their features.

Here is an original prompt for a background concept:

Use the attached approved green-bottle photograph as the product reference. Create a calm tabletop scene with warm off-white surroundings and soft light from the left. Leave open space to the right for text added later. Keep the bottle upright and preserve its visible shape, cap and label as closely as possible. Do not invent accessories, extra printing or a new product variant. This is a background concept; the source photo remains the authority for the product.

The last sentence is a review rule for you as much as an instruction to the model. Save the generated result separately. Never overwrite the approved source.

If the result changes the bottle, move to a composition workflow: generate or photograph the empty background, then place the original product cutout above it. Match its scale, contact shadow and light direction without repainting the label.

Review details before judging the mood

Put the reference and candidate side by side at the same apparent product size. Check the most important region first. For a skincare jar, that could be the label and lid. For a backpack, it could be the straps, seams and pockets.

This is a practical review checklist, not a scientific measurement of fidelity. If exact colour is commercially important, involve the person responsible for product imagery and use an appropriate colour-managed workflow.

Repair the right layer

For a controlled composite, place the actual product photo above the new background and hide unwanted surroundings with a layer mask. Adobe documents masks that reveal or hide portions of a layer; keep the product's original image available underneath your editing decisions. Photoshop layer-mask instructions

A missing letter is sometimes easy to fix; a different bottle geometry is not. Do not cover a large product mismatch with a logo and call it accurate.

For a label that should remain flat and front-facing, an editor may be able to place the exact approved artwork over the correct product photograph. For a curved or angled package, the artwork needs realistic perspective and surface fitting. If that work is beyond the team’s skill, use the original photograph or reshoot.

For a wrong product colour, first ask whether the light in the scene is causing a colour cast. A warm light can affect the whole scene. Recolouring only the product may make it look pasted in. Adjust carefully and compare to the real item; do not simply pick the most attractive green.

For a wrong shape, return to the source. Repeatedly asking the model to “fix everything but keep everything” can create a chain of new changes. A clean restart with fewer ambitions may be easier to review.

Keep text and evidence outside the generated scene

Add prices, offer dates, specifications and calls to action as editable text layers. This avoids repainting the entire image when an offer changes. It also lets you proofread the exact words independently.

Do not generate a fake review, invented award badge or before-and-after result. A fictional scenario should be visibly presented as an illustration when viewers could otherwise read it as an actual event. Follow the publishing platform’s current requirements for AI-generated media and preserve source provenance.

Run the same acceptance protocol on every candidate

Write the pass conditions before generating. For this fictional bottle task, the correct variant, readable approved label, cap, silhouette and absence of invented claims are mandatory. A beautiful background cannot compensate for a failed mandatory check. If a detail cannot be verified, mark it “not verified” and keep the candidate out of the accepted total.

Use a small, named batch with a spending limit you choose. The number of attempts is a planning input, not a recommendation or a promise of success. Record the actual model identifier, app or interface, date, prompt version, source-reference version and output settings available to you. “Nano Banana” alone is not a complete production record: Google's documentation covers distinct models and capabilities. Check the current official image-generation documentation for the model you actually use; it was reviewed for this guide on September 20, 2026.

Follow the same sequence for each output:

1. Register it. Save a unique filename and preserve the source, prompt and unedited candidate. Count rejected attempts too.
2. Compare product facts. Match apparent product size, then inspect the label, closure, edges and parts against the approved reference. Read every visible character; do not accept plausible-looking text.
3. Check the whole scene. Inspect reflections, shadows, hands, accessories and scale for invented features or misleading implications.
4. Review the destination. Look at the actual crop and export size, not only a large preview. Confirm editable text, permissions and campaign claims.
5. Record a decision. Choose accept, reject, or hold for a specific repair or missing verification. Give the reason and reviewer, rather than a vague quality score.
6. Retest revisions. Save a new version and repeat the complete checklist. Repairing a label does not exempt the cap or colour from another check.

Use the same source and criteria when comparing two workflows. Do not choose a looser standard for the more expensive output because you have already paid for it. For repeated production, review another representative batch when the model, source product, prompt or editing method changes. Keep the batches separate so an old successful image does not stand in for today's checks. This is an operational review process, not a controlled model benchmark.

Fictional acceptance table: six imagined candidates

The following decisions are invented to demonstrate the worksheet. No images were generated or tested for this table. They do not show a model's measured accuracy or a likely acceptance rate. “Pass” below means the stated hypothetical checks are assumed to have passed.

A hold is not a pass. If C06 is later cleared, record the evidence and decision date; only then update the accepted count. If C03 is repaired, create a revision record and inspect the entire revision. Do not count the original and revised file as two accepted deliverables when they are versions of the same final image.

The useful output of this exercise is a rejection reason you can act on. Three label failures suggest changing the editing approach, not merely making another mood variation. A reference photograph or original-product composite can be the right final choice.

Calculate cost per accepted output, not per attempt

Here is a separate, fictional budgeting example using the table's assumed two accepted outputs. The numbers are illustrative inputs, not Google prices, actual charges or a test result.

The ₹300 is a valuation of time unless someone was actually paid that amount; do not present it as a cash expense automatically. Track cash spend and time separately, then state which you include in the combined figure. Add actual retouching, reference-shoot or licensed-asset costs if they belong to your comparison, avoiding double counting.

Include failed attempts and review of rejected candidates in the batch totals. When comparing against a photo or cutout workflow, use the same deliverable standard and cost boundary. The lower generation price alone cannot tell you which route is cheaper to finish.

If the accepted count is zero, cost per accepted output is undefined, not ₹0. Report “no accepted outputs,” the total spend and time, and the next decision: stop, revise the method, or use the approved photograph. Do not divide by one to make a number appear.

Decide what “approved” means

Use three separate sign-offs:

1. Product accuracy: the item matches the approved variant.
2. Creative quality: the image works at its intended size and crop.
3. Publication readiness: text, claims, permissions and destination are correct.

A designer can approve the layout while the product owner rejects the label. Keep those decisions separate in the worksheet.

If you cannot verify the generated product, do not use it to sell that specific item. Use a real cutout, a real photograph or a clearly illustrative scene without product-detail claims. The goal is not to prove that AI can do every shot. It is to finish the campaign with imagery you can stand behind.

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: