GPTWala

Tag: ecommerce images

  • Common AI Product Photography Mistakes: Troubleshooting and Rejection Checklist

    Reviewer tracing an AI product-image defect from the source photo to the final export
    Find the first file where product truth fails; repair that stage or recapture the missing evidence. Original GPTWala diagnostic illustration using one fictional, unbranded product; not a client result, tool test or platform interface.

    Reviewed and updated: 12 August 2026

    Editorial disclosure: GPTWala did not run a product shoot, AI editor, marketplace submission or controlled accuracy test for this article. The diagnostic system is editorial guidance based on current official sources and production reasoning. Product owners must compare every candidate with the exact physical SKU and verified product records.

    The safest way to fix an AI product-photo mistake is to find the first file where the product becomes wrong. Compare the source, mask, generated candidate, composite and export in order. Repair only that failing stage. Reject and recapture when identity, quantity, label, geometry, material, scale or another buying-relevant fact cannot be verified. Do not keep regenerating until a plausible image hides the defect.

    Table of contents

    1. What counts as an AI product-photo mistake?
    2. Quarantine the image before troubleshooting
    3. Find the first wrong file
    4. Use the complete troubleshooting table
    5. Diagnose identity, text and quantity failures
    6. Diagnose shape, colour and material drift
    7. Diagnose edges, shadows and scale
    8. Diagnose file, channel and hand-off failures
    9. Choose a safe repair level
    10. Know when to stop and recapture
    11. Apply the method to Indian product businesses
    12. Prevent repeat failures without building more bureaucracy
    13. Run the rejection checklist
    14. Frequently asked questions

    What counts as an AI product-photo mistake?

    An AI product image has failed when it is unusable for its intended job, even if it looks polished. There are four different failure types:

    Failure type The image may look like The real problem Appropriate response
    Product-truth failure Attractive and believable A buyer-relevant product or offer fact changed Reject; restore verified evidence or recapture
    Presentation failure Rough, cut out or poorly grounded Product may still be correct, but the image distracts or confuses Repair the background, edge, shadow or crop without touching the SKU
    Destination failure Correct on the editing screen The crop, overlay, file, metadata or current channel rule fails after export Re-export from the approved master and verify the destination
    Process failure Each file looks acceptable alone Wrong SKU mapping, inconsistent batch style or unclear approval lets the wrong asset travel Quarantine the batch; correct mapping, ownership or hand-off

    The most dangerous mistake is not always the most obvious. A rough shadow is visible and usually repairable. A subtly changed valve port, sari border, stone setting or pack quantity can look professional while describing the wrong product.

    This guide owns symptom-led diagnosis and the fix-or-reject decision. It does not repeat the full phone-to-approved image workflow, the one-phone-photo tutorial, the background-generation workflow or the broader AI image accuracy governance system. Use those pages for their respective jobs.

    A mask is not a product lock

    Selecting only the background does not prove that the product will remain untouched. OpenAI’s current image-editing guidance says selections are not always precise and an edit may extend beyond the targeted area. Google’s current Product Studio guidance calls its generative features experimental and warns that unexpected outputs may occur. Those are useful operating cautions, not evidence that every edit will fail.

    Assume every generated candidate is unapproved until it has passed a comparison with the exact source and the physical or recorded product truth. A prompt can state a constraint; it cannot sign off the output.

    Quarantine the image before troubleshooting

    When someone notices a defect, stop the candidate from moving into a catalogue, ad folder or seller upload. Do not overwrite the approved source or rename the faulty export as final.

    Record these eight facts before editing again:

    1. exact SKU, child variant and offer quantity;
    2. intended image role: proof, main, detail, context, ad or internal concept;
    3. destination and current specification owner;
    4. source file used, including date or version;
    5. first visible symptom;
    6. first file in which the symptom appears;
    7. product record, physical sample or source view used to verify it; and
    8. decision: narrow repair, recapture, change method, specialist review or reject.

    Call this a defect card. It need not be a new software system. One row in the existing production register is enough.

    Describe the symptom without guessing the cause

    Write “the right handle disappears at the rear edge” before writing “bad prompt.” Write “the blue child SKU is attached to the green-variant file” before writing “AI colour problem.” The first statement can be checked. The second can send the team to the wrong repair.

    Avoid vague diagnoses such as:

    • “looks fake”;
    • “AI issue”;
    • “make premium”;
    • “colour is off” without naming the reference and viewing condition; or
    • “marketplace rejected” without recording the actual account message and submitted file.

    A precise symptom narrows the investigation. It also prevents a team from regenerating the product when the real fault is a crop preset, a mislabelled source or a compressed export.

    Find the first wrong file

    Follow the files in production order:

    Verified source → selection or mask → generated candidate → composite → approved master → destination export

    Open them side by side at useful magnification. Ask one question at every step: Is the named defect already present here?

    First wrong stage What it usually means First safe action
    Verified source The camera did not capture the field, the wrong SKU was photographed, or the record is incomplete Stop editing; correct the SKU mapping or recapture
    Selection or mask Fine edges, holes, transparent areas or gaps were included/excluded incorrectly Rebuild a smaller, cleaner non-destructive selection
    Generated candidate The editor altered protected pixels, inferred an unseen detail or introduced an object Reject candidate; constrain the edit or retain the real product layer
    Composite Light, perspective, contact, scale or occlusion no longer agrees Rebuild the composite with measured geometry and a real retained product layer
    Approved master Approval was attached to the wrong version or an unverified repair was flattened in Revoke approval; return to the last verified file
    Destination export Crop, resize, colour conversion, metadata handling or overlay changed the approved asset Re-export from the approved master; do not regenerate

    This “first wrong file” method matters because late-stage repairs can conceal an early truth failure. If the source never shows the back label, sharpening the final image cannot recover it. If the approved master is correct but a square preset cuts off the handle, a new AI image is unnecessary.

    Run one diagnostic check, not five speculative edits

    Choose the smallest test that can confirm or reject the suspected cause:

    • toggle the candidate over the source at 50% opacity;
    • place matching landmarks on silhouette, holes, seams or corners;
    • compare an exact label crop with the verified artwork or source photo;
    • count components and included items;
    • view source and candidate under the same colour-managed conditions;
    • disable the generated background and inspect the product edge;
    • compare the approved master with the delivered export; or
    • open the production register and verify the SKU-to-file relationship.

    If the check does not isolate the problem, return to the symptom. Do not compensate by adding more prompt adjectives.

    Diagnostic flow from source photo through mask, AI candidate, composite and export to the safest repair

    Fix the earliest wrong stage; do not conceal it downstream. Original GPTWala diagnostic flow with native labels; it reports no provider score, model test or marketplace result.

    AI product photography mistakes: symptoms, causes, fixes and stop rules

    Use this table as triage. “Likely cause” is a hypothesis to test, not a diagnosis made from appearance alone.

    Symptom Likely cause Diagnostic check Safe fix Stop or recapture when
    Wrong product or neighbouring variant Wrong source, filename or SKU mapping Match physical item, SKU record and source identifier Correct mapping; restart from the verified source Exact variant cannot be established
    Label, logo or printed text is garbled Generative reconstruction, low-resolution source or aggressive enhancement Compare characters, line breaks, placement and legal/product fields with verified artwork and real pack Restore exact approved artwork or untouched real label layer Source/artwork is missing, outdated or unreadable
    Pack count or included accessory changes Model inferred a set or styling prop; offer record was vague Count every sale unit and component against the offer record Remove non-included props non-generatively; rebuild from the exact quantity source It is unclear what the customer receives
    Product gains or loses a part Occlusion, incomplete source pack, mask error or generative completion Compare front, back and detail views; trace the part into the mask Restore verified pixels; repair mask narrowly Part is not visible in any source or affects function/safety
    Silhouette, port, seam or construction drifts Broad edit changed protected geometry Overlay source and candidate; pin landmark coordinates Use retained real product pixels or revert the generation Geometry cannot be restored without invention
    Product looks stretched or tilted Perspective correction, resize or compositing mismatch Compare corner/axis landmarks and source aspect ratio Re-transform from the original with proportions locked Dimensions or fit would be materially misrepresented
    Colour variant shifts Mixed light, auto correction, background influence, colour conversion or generative relight Compare with physical item and neutral reference; inspect the approved master/export path Correct conservatively from a verified reference; publish multiple truthful views if appearance varies No trustworthy colour reference exists
    Matte becomes glossy, metal becomes plastic, weave disappears Smoothing, relighting, denoising or invented material Compare highlight shape and microtexture at full resolution Restore source texture; reduce the edit to the surrounding area Material/finish cannot be verified after repair
    Pattern, print or texture repeats incorrectly Generative fill tiled or reconstructed detail Align motifs, border sequence, grain and intentional irregularity Restore the real product layer or exact verified texture region Pattern is a selling feature and source evidence is incomplete
    Jewellery stone, prong, clasp or link changes Fine repeated geometry was generated or erased Count stones/settings; compare macro and construction views Reject candidate; use real macro or jewellery specialist workflow One buying-relevant element differs or a mark is unclear
    Apparel fit, drape, neckline or border changes Garment was re-generated on a model; source does not prove worn behaviour Compare flat, mannequin and measured references; inspect seam/border map Use retained garment layer or real model/mannequin capture Fit, coverage, fall or construction is a purchase decision
    Halo, missing edge or jagged cutout Mask includes background or removes fine/transparent detail View on black, white and mid-grey; toggle mask edge Rebuild mask from source; use manual or specialist cutout Edge cannot be separated without inventing fibres, chain or transparency
    Product floats or shadow points the wrong way Contact point, light direction or surface plane mismatch Draw baseline and light direction; inspect gap at contact edge Rebuild a subtle physically consistent shadow outside the product Product scale/contact cannot be verified in the scene
    Product appears too large or small in context Unmeasured scene, generated hand/model or lens/perspective mismatch Compare recorded dimensions with a known plane or reference object Rebuild at measured scale; label dimensions accurately Context determines fit, clearance or safe use and measurement is absent
    Context implies an unsupported use Prompt created installation, ingredient, compatibility or performance meaning Ask what claim a reasonable buyer could infer; compare product records Choose a neutral context or a verified real use case Use, compatibility, safety or performance is not documented
    Crop hides a handle, connector, border or pack edge Destination template or subject detection cropped the product Compare approved master and export with safe-area overlay Re-export with a product-specific crop Required identifying or functional feature will not fit the format
    Text overlay becomes part of the product offer Promotional badge, price or claim overlaps or appears printed on pack Compare clean master and destination creative; read the full message Keep clean commerce master; add only reviewed native overlay for an allowed role Destination forbids overlay or claim is unsupported
    Upscale looks sharp but creates false microdetail Generative upscale or sharpening fabricated texture/characters Compare pixels with the highest-quality real source, not only the low-res version Use a better source or conservative non-generative resize Detail is needed to verify label, finish, setting or construction
    AI/provenance metadata disappears Export, conversion, CDN or download path stripped metadata Inspect the actual delivered file with a metadata reader Re-export through a tested path; retain original and provenance record Destination requires metadata and preservation cannot be confirmed
    Correct image is attached to the wrong listing Manual copy, reused folder, ambiguous filename or variant merge Reconcile file ID, SKU, product record and destination item ID Correct the mapping and review affected neighbouring records Scope of the mapping error is unknown; quarantine the batch
    Batch style changes from one SKU to the next Prompts, templates, reviewers or source angles vary View contact sheet grouped by visual family while keeping SKU truth cards open Correct presentation controls in a small batch “Consistency” repair would change a real variant field
    Final file differs from the approved master Wrong version, compression, colour conversion or post-approval edit Hash/version check where available; visually compare exact delivered file Replace with a fresh derivative from the approved master Approval trail cannot identify the released source

    Do not turn this table into a blind automation

    The table helps a reviewer choose the next check. It cannot see the physical SKU, know the seller’s offer or decide whether a material difference matters. A reviewer with product authority must make the final call.

    Diagnose identity, text and quantity failures

    Identity, text and quantity failures are automatic commercial risks because they can change what the customer believes they will receive.

    Wrong SKU is a mapping problem until proven otherwise

    Before blaming the model, inspect the folder and register. Similar variants are easily confused: two Morbi tile finishes, adjacent bottle sizes, right- and left-hand machine components, a necklace sold with or without earrings, or a sari design in two border colours.

    Verify:

    • physical sample or authorised source ID;
    • exact child SKU and revision;
    • colour, size, finish and configuration;
    • pack quantity and included components;
    • source date; and
    • destination item ID.

    If the source belongs to another variant, no prompt can repair the mapping. Start again with the correct record.

    Restore text; do not rewrite it from memory

    Labels can contain identity, ingredients, capacity, warnings, directions, certification references, manufacturer information and other important fields. A visually plausible replacement is not acceptable.

    Use one of these routes:

    1. retain the exact real label pixels when they are clean and legible;
    2. place authorised current artwork at the verified angle and dimensions, then obtain owner approval; or
    3. recapture the pack or label.

    Do not ask a generative model to “make the text readable.” Do not reconstruct blurred characters from memory. Do not borrow artwork from a related size or market. If the current artwork is disputed, the content owner—not the image operator—must resolve it.

    Count the offer twice

    Count the physical items in the source and count the items in the final candidate. Then compare both with the actual offer record.

    A styling bowl beside a spice pack can look included. A generated necklace set can acquire a second bangle. A B2B component image can show four pieces although the quote is per piece. A “pair” can accidentally become one item through cropping.

    Where the offer is ambiguous, stop. Fix the commercial record before making the image.

    Diagnose shape, colour and material drift

    These errors often survive a quick review because the candidate looks believable. Inspect with the physical product nearby whenever possible.

    Geometry: use landmarks, not overall resemblance

    Choose points that should not move:

    • outer corners and silhouette breaks;
    • hole, port, handle and fastener centres;
    • neckline, seam, hem and border intersections;
    • clasp, prong, hinge and joint positions;
    • cap, shoulder, base and label boundaries; and
    • intentional gaps or negative spaces.

    Overlay the candidate on the source and toggle visibility. Small camera changes can prevent perfect pixel alignment, so the purpose is not to manufacture a numeric accuracy score. It is to expose a changed construction, proportion or missing part.

    Recent research still treats fine-grained product identity preservation—including branding and text—as a hard image-editing problem. The 2026 ProductConsistency paper is a preprint, not a commercial tool guarantee, but its problem framing supports the conservative rule: plausible resemblance is not proof of exact product preservation.

    Colour: trace the whole path

    Colour can change at capture, edit, compositing, export or display. Diagnose in order:

    1. Was the source captured under mixed or strongly coloured light?
    2. Is there a neutral reference or the physical product for comparison?
    3. Did the background create a visual colour contrast?
    4. Did the AI relight or “enhance” the product?
    5. Did the export change colour space or profile?
    6. Does the delivered file differ from the approved master?

    Do not promise that every viewer will see an exact screen match. Preserve the real variant, avoid dramatic colour grading and give multiple truthful views when a finish changes with angle or light.

    Material: keep the cues that make it identifiable

    Material truth often lives in small cues: weave, grain, pores, brushed lines, edge highlights, translucency, uneven handmade texture or surface reflection. Removing all “imperfections” can remove the product itself.

    Reject an edit that:

    • turns brushed metal into mirror chrome;
    • smooths handloom weave into synthetic-looking fabric;
    • makes glazed tile appear matte or vice versa;
    • converts translucent packaging into opaque plastic;
    • invents uniform sparkle across jewellery; or
    • fills wood, leather or stone with a repeated synthetic texture.

    If those cues were not captured, recapture under better light. An upscale cannot reveal real detail that the camera never recorded.

    Diagnose edges, shadows and scale

    These are presentation problems until they begin changing product meaning.

    Check edges on three backgrounds

    Place the cutout against black, white and mid-grey. This reveals white fringes, dark contamination, missing translucent detail and over-feathered edges. Inspect at normal page size and at magnification.

    Repair the selection, not the product. Fine apparel fibres, glass edges, jewellery chains, handles, holes and open metalwork may need a manual path, channel-based mask, specialist retouch or a better source. If the edge cannot be separated without reconstructing the product, stop and recapture.

    For the full process of creating a new setting while protecting product pixels, use the AI product-background generation guide. This troubleshooting page only diagnoses the fault.

    Ground the product before beautifying the scene

    A grounded product needs agreement between contact, surface plane, perspective, shadow direction and light. Use a simple diagnostic:

    • draw the contact baseline;
    • mark the dominant light direction;
    • identify the surface plane;
    • check whether the shadow begins where the product touches it; and
    • ask whether the shadow softness fits the apparent light size and distance.

    If the scene is too complex to solve without changing the product, simplify it. A neutral background with a quiet, physically plausible shadow is safer than an impressive room that makes the product float.

    Measure scale when context influences purchase

    A generated hand, shelf, room or model can change apparent size. Record product dimensions first. Then check the object’s placement on a known plane or use a real measured reference outside the clean image.

    Use real capture when context answers a fit, clearance, installation or safety question: garment fit, jewellery fall, furniture proportions, machine clearances, connector placement or an item worn near the face/body. A generated context can illustrate an idea; it should not become unverified measurement evidence.

    Diagnose file, channel and hand-off failures

    The product can survive the AI edit and still fail after approval.

    Reopen the delivered file

    Inspect the exact file that a website, feed, dealer or marketplace will receive—not only the editor canvas. Check:

    • product and offer still match;
    • crop retains the whole required view;
    • small text and detail remain legible where needed;
    • colour and transparency behave as expected;
    • filename/version maps to the correct SKU;
    • overlays are allowed and supported;
    • file format and size match the current destination; and
    • required provenance metadata remains in the delivered asset.

    Google Merchant Center’s current main-image guidance asks for the actual, correct product and variant, including colour, pattern and material, and restricts placeholders and promotional overlays. Its current AI-content guidance says applicable generative-AI product images submitted in specified image attributes should retain the named IPTC digital-source metadata. Those are Google-specific requirements; use the product-image rules by channel for a broader destination check.

    Do not assume that an editor export, WebP conversion, WordPress optimisation plugin, CDN or platform download has preserved metadata. Test the real delivery path. C2PA’s own explainer also cautions that provenance can be incomplete or removed and that provenance alone does not establish whether content is true. Keep product review and file provenance as separate checks.

    Treat “rejected by platform” as an observed event, not a diagnosis

    Record:

    • the exact submitted file;
    • seller account, category and destination;
    • submission time;
    • actual status or message;
    • any human-support response; and
    • the next controlled change.

    Do not invent a reason from a generic article. Do not claim that a file is “marketplace-approved” because it looks compliant. Rules, enforcement and account states can differ and change.

    Misleading can happen by implication

    The Advertising Standards Council of India’s current code says advertisements should be truthful and honest and that visual presentation should not mislead through implication, omission, ambiguity or exaggeration. ASCI is a self-regulatory body, and this article is not legal advice. The practical lesson is simple: a false impression can come from scale, context, included props or a perfected material—not only from written copy.

    Choose the safe repair level

    Use the lowest repair level that can restore a verified result.

    Level Action Appropriate when Never use it to
    0 — Mapping correction Attach the right approved file to the right SKU/destination Image is correct; relationship is wrong Pretend a neighbouring variant is acceptable
    1 — Re-export Create a fresh crop/format/size from the approved master Failure appears only after export Rebuild missing product detail
    2 — Narrow presentation repair Correct mask, dust, outside background or physically consistent shadow Product pixels and truth fields remain verified Alter label, construction, material or quantity
    3 — Restore verified product evidence Return the real product layer or authorised artwork Generation damaged a known field but exact evidence exists Invent unseen sides or characters
    4 — Recapture Photograph the exact item, angle, label, texture, scale or part again Source evidence is missing or technically unusable Avoid resolving an uncertain SKU/offer record
    5 — Change method or specialist Use real studio, hybrid composite, retoucher or category specialist Repeated defects affect high-risk detail Turn an unverified candidate into proof
    6 — Stop/reject Remove asset from production Truth, rights, safety or offer cannot be verified Keep a plausible image because of deadline pressure

    One controlled repair is better than a chain of untracked regenerations. If the same locked field fails again, escalate the method. Repetition is evidence that the current lane is a poor fit for that product, not an invitation to lower the approval standard.

    When to stop AI and recapture the product

    Recapture is mandatory when the source cannot prove a buying-relevant fact and no exact verified asset can restore it.

    Stop and use real capture when:

    • exact SKU or child variant is uncertain;
    • label, legal text, warning, mark or identifier is unreadable;
    • a hidden side contains ports, seams, fasteners, ingredients, settings or accessories that matter;
    • colour/finish is important and no trustworthy reference exists;
    • count, quantity or included components are disputed;
    • scale, fit, drape, clearance or installation is a buying decision;
    • jewellery settings, hallmark area or fine construction cannot be checked;
    • an AI enhancement has invented microdetail;
    • transparent, reflective or fine-edged material cannot be separated reliably;
    • rights to the source, artwork, model or reference are unclear;
    • a safety, compatibility, certification or performance impression cannot be supported; or
    • two controlled attempts repeat the same truth failure.

    The “two attempts” point is a practical escalation rule, not a universal accuracy statistic. A critical identity failure can require stopping after the first candidate. A harmless crop adjustment may take more than two non-generative exports.

    Automatic-reject fields

    Reject immediately if the final asset changes or leaves unresolved:

    • product identity or variant;
    • offer quantity or included component;
    • label, logo, mark or buying-relevant text;
    • silhouette, construction, fit or functional geometry;
    • material, finish, pattern or meaningful colour;
    • product scale where context affects the decision;
    • compatibility, safety, use or performance implication;
    • rights or consent; or
    • final SKU-to-file mapping.

    The product-accuracy audit for AI images remains the owner of the full governance and approval record. This page tells the reviewer what to do after a symptom appears.

    Troubleshooting examples for Indian product businesses

    These are fictional operating examples, not client results, city-wide claims or tested tool outcomes.

    Surat apparel seller: the sari border changes on a model

    Symptom: motifs near the pallu repeat differently and the border becomes narrower.

    First wrong file: the generated model candidate; the flat source and mask are correct.

    Diagnostic: compare the border sequence, seam intersections and pallu map with the real sari. Check whether the garment was re-generated rather than retained.

    Decision: reject. Use the real garment layer, a controlled mannequin composite or a real model shoot. Fit and drape need the specialist AI model-photo workflow for apparel; prompt repetition is not a safe repair.

    Jaipur jewellery retailer: an earring gains a stone

    Symptom: the product still looks symmetrical, but one accent stone and two prongs differ.

    First wrong file: the generated candidate.

    Diagnostic: count each stone and setting against a real macro of both actual earrings. Check backs and offer quantity separately.

    Decision: reject the candidate. Restore real pixels or recapture. Use the AI jewellery photography truth checklist for stone, setting, clasp, reflection and hallmark-specific review.

    Rajkot component manufacturer: the threaded port is softened

    Symptom: an internal thread looks smooth and the hole diameter appears larger.

    First wrong file: an aggressive cleanup/upscale.

    Diagnostic: compare the original macro and engineering/product record; inspect whether the feature is necessary for identification or fit.

    Decision: stop AI enhancement. Recapture the port or supply an approved technical/detail photograph. Do not use a generated thread as compatibility evidence.

    Morbi tile wholesaler: two finishes become one

    Symptom: matte and satin variants look nearly identical after background and colour standardisation.

    First wrong file: the batch composite; the source files preserve the difference.

    Diagnostic: compare highlight width, surface texture and child-SKU mapping under the same viewing conditions.

    Decision: restore each real surface and create separate visual-family settings if required. Consistency should standardise presentation, not erase the finish a buyer orders.

    Packaged-goods retailer: the front label is “cleaned up”

    Symptom: the brand looks correct at a glance, but one quantity line and two characters differ.

    First wrong file: the generated enhancement.

    Diagnostic: compare with current authorised artwork and the exact physical pack; verify the pack size and market version.

    Decision: reject. Use real label pixels, exact authorised artwork with owner sign-off, or a new capture. Never recreate packaging text from memory.

    Multi-SKU wholesaler: the correct image reaches the wrong row

    Symptom: the image itself passes review, but a six-hole part appears against the four-hole SKU.

    First wrong stage: catalogue mapping after approval.

    Diagnostic: reconcile asset ID, child SKU, product record and destination item ID; inspect adjacent rows for the same copy error.

    Decision: quarantine the affected batch, repair the mapping and re-review the release manifest. Use the AI catalogue photography system for manufacturers and wholesalers to prevent recurrence.

    Prevent repeat failures with a small defect log

    A defect log should accelerate production, not become a new project. Add one row only when a candidate fails or requires a consequential repair.

    Field Example value
    Asset/SKU Fictional SKU MUG-TEAL-02
    Symptom Right handle gap filled
    First wrong file Generated candidate v03
    Suspected layer Selection/reference
    Diagnostic check Source/candidate overlay; black-background edge check
    Safe action Rebuild mask; retain real handle pixels
    Result Candidate rejected; v04 sent to review
    Reviewer/date Named product owner / date

    Use short reason codes to make patterns visible:

    Code Meaning Example
    ID01 Identity or variant Wrong child SKU
    OF01 Offer, quantity or text Extra accessory; pack count changed
    GE01 Geometry or construction Missing handle; changed seam
    MA01 Material, colour or pattern Matte became glossy
    CT01 Context, scale or claim Product floats; unsupported installed use
    PL01 Platform/destination Disallowed overlay or current rule conflict
    EX01 Export/delivery Crop, compression, colour or metadata loss
    RT01 Rights/consent Source or reference permission unresolved

    Review the log after a meaningful batch, not after every pixel change. If the same code repeats for one product family, alter the source pack, template, method or review gate. Do not interpret a small internal log as a model-wide accuracy benchmark.

    Contact sheet illustrating wrong label, geometry, material, floating, scale and export defects

    Illustrative defects built manually from one locked fictional product for training; not observed results or a model comparison.

    Final AI product image rejection checklist

    Use this on the exact delivered file. One “no” in a critical field keeps the asset out of production.

    Identity and offer

    • [ ] Exact child SKU and revision are confirmed.
    • [ ] Colour, size, finish and configuration match.
    • [ ] Pack quantity and every included component match the offer.
    • [ ] No prop or context object appears included by mistake.
    • [ ] Label, logo, mark and product text match verified evidence.

    Product construction and appearance

    • [ ] Silhouette, dimensions and proportions are not stretched.
    • [ ] Ports, holes, handles, seams, fasteners, settings, joints and accessories are complete.
    • [ ] Pattern, border, texture, grain and intentional irregularity remain real.
    • [ ] Material, finish, colour and reflection cues remain truthful.
    • [ ] No generated detail is being used as proof.

    Presentation and context

    • [ ] Edge is clean on light, dark and mid-tone backgrounds.
    • [ ] Product contact, perspective, light and shadow agree.
    • [ ] Context does not imply unsupported size, fit, installation, safety, compatibility or performance.
    • [ ] Crop preserves all features needed for this image role.
    • [ ] Any native text overlay is accurate, approved and allowed for the destination.

    File and release

    • [ ] Final delivered file matches the approved master.
    • [ ] Filename, asset ID and destination item map to the correct SKU.
    • [ ] Current account/category/platform requirements were checked at publication time.
    • [ ] Required AI/provenance metadata is present in the actual delivered file where applicable.
    • [ ] Rights, model consent, artwork authority and reviewer sign-off are recorded.

    Record approved, rework, real capture required or rejected. “Looks fine” is not a release status.

    Turn fewer rejected images into a stronger online system

    Troubleshooting is useful when it gets verified product content moving again. It should not become endless image polishing.

    The GPTWala workshop connects this AI Content Creation work to the DAA path: Digital Presence → AI Content Creation → ₹100/day WhatsApp ads. The workshop is educational; it does not promise enquiries, sales, earnings or return on ad spend. Advertising claims, targeting, landing pages and follow-up still need their own decisions and controls.

    See the GPTWala workshop and decide whether the DAA framework fits your product business.

    Frequently asked questions

    Why does AI keep changing product labels and logos?

    Generative editors may reconstruct small or complex text instead of preserving exact pixels, especially when the source is low resolution or the edit touches the product. Compare the candidate with current authorised artwork and the physical pack. Restore the real label layer or recapture it; never rebuild buying-relevant text from memory.

    Why does my product change colour or shape after a background edit?

    The selection may include product pixels, the editor may relight or re-generate the object, or the final export may change colour or proportions. Find the first wrong file, then test the mask, source overlay and approved-master/export path. If no verified colour or geometry reference exists, recapture.

    Should I fix an AI product image or generate it again?

    Use a narrow fix when the source is verified and the defect is limited to presentation or export. Regeneration is not automatically safer. If identity, text, quantity, construction or material changed, restore real evidence, change the method or recapture. Reject repeated truth failures.

    What should I do when AI removes a handle, clasp or small part?

    Check whether the part is present in the source and mask. If verified real pixels exist, rebuild the selection and restore them. If the part is hidden, blurred or absent from every source, photograph it. Do not ask AI to guess functional construction.

    How do I fix a floating AI product photo?

    Check the contact baseline, surface plane, perspective, light direction and shadow origin. Rebuild only the scene and shadow around a retained real product layer. If product size or placement cannot be measured, simplify the background or use a real contextual capture.

    Can a better prompt guarantee product accuracy?

    No. A constraint prompt can reduce ambiguity, but it cannot verify the result or make an unseen detail true. Use exact sources, narrow edits and a human comparison. The product owner—not the prompt—approves identity and offer fields.

    Does AI metadata prove an image is accurate?

    No. Provenance metadata can help describe an asset’s history, but it may be incomplete or removed and does not prove the depicted product is true. Check both the delivered file’s required metadata and the product itself against verified evidence.

    When is one phone photo not enough?

    One view is not enough when the missing side contains a label, pattern, component, mark, clasp, seam, port, texture or dimension needed for purchase or review. Capture additional real views. The single-phone-photo tutorial is for controlled presentation candidates, not invention of unseen product truth.

    When should I use a photographer or specialist instead of AI?

    Use a photographer, retoucher or category specialist when accurate colour, fine construction, reflective/transparent material, apparel fit, jewellery detail, regulated information, installation or high-value proof cannot be captured and verified in the AI lane. The AI versus studio versus hybrid guide helps select the method.

    Sources and review method

    Reviewed 12 August 2026. Official sources were used for named editor limitations, Google product-image and AI-metadata requirements, Indian advertising context, structured-data implementation and provenance cautions. One current research preprint is used only to support the continuing difficulty of exact product-identity preservation, not as a tested tool result. Operational tables, reason codes, repair levels and Indian examples are original GPTWala editorial guidance. Recheck platform- and account-sensitive claims within 24 hours of publication.

  • AI Product Photography Prompt Pack That Protects Product Truth

    Fictional terracotta jar shown as a neutral source reference and in a warm contextual scene, separated by four prompt-layer cards
    AI-generated editorial illustration using a fictional, unbranded reference image. It is not a merchant result, physical-SKU test or product-accuracy benchmark.

    Reviewed and updated: 12 August 2026

    Template status: every prompt on this page is a tool-neutral template to verify with your own SKU. GPTWala has not labelled these templates “tested” because a dated, controlled exact-SKU prompt test has not yet been completed. A prompt can direct an edit; it cannot certify the output.

    A safer AI product photography prompt has four layers: source truth, permitted change, scene specification, and negative plus acceptance conditions. Attach photographs of the exact SKU and name the details that must not change. Words such as “photorealistic” or “premium” are not evidence. Negative instructions reduce ambiguity, but they cannot guarantee fidelity; compare every output with the real product and reject any material change.

    Table of contents

    1. The safest prompt formula
    2. Make a product truth card
    3. Catalogue and main-image prompts
    4. Lifestyle and additional-image prompts
    5. Specialist product prompts
    6. Ad-creative prompt
    7. Prompt repair ladder
    8. How to test prompts
    9. Indian business adaptations
    10. What prompts cannot solve
    11. FAQs

    The safest AI product photography prompt formula

    Tell the tool what is true before telling it what to create. A commercial product-image prompt should answer four questions in this order:

    Layer Question it answers What belongs here
    1. Source truth Which exact sale item is authoritative? SKU, variant, supplied views, locked visible attributes and verified dimensions
    2. Permitted change What is the tool allowed to edit? Background, selected region, canvas, surrounding light or other narrow change
    3. Scene and output What useful image should be made? Image role, destination, setting, viewpoint, crop, contact shadow and aspect ratio
    4. Negative + acceptance conditions What causes rejection? Prohibited additions and a clear stop rule for any change to product or offer truth

    Four prompt layers moving from source truth to permitted change, scene specification and rejection conditions

    Original GPTWala prompt-anatomy diagram. A prompt narrows the edit boundary; it does not guarantee that a model will preserve the product.

    Copy this modular master prompt and replace every field in square brackets:

    SOURCE TRUTH
Use the attached photographs of the exact [SKU and variant] as the only source of
product identity. The [front / 45-degree / back / label / detail] references all show
the same physical item. Locked attributes: [silhouette and proportions], [colour],
[material and finish], [pattern], [label/logo/text], [quantity and included parts],
and [verified dimensions or supplied scale cue].

PERMITTED CHANGE
Change only [background / canvas outside the product / lighting around the product /
selected region]. Keep the supplied product layer intact. Do not redraw, recolour,
relabel, resize, beautify, add to, or remove any part of the product.

SCENE AND OUTPUT
Create a [main catalogue / additional / lifestyle / dealer-detail / ad] image for
[destination and audience]. Show [setting and surface] from [viewpoint], with
[lighting direction], a physically plausible [contact shadow/reflection], [crop],
and [aspect ratio]. Keep props secondary and scale believable.

NEGATIVE + ACCEPTANCE CONDITIONS
No extra product units, accessories, text, logos, badges, hands, claims or offer
elements unless they are supplied and verified. Reject the result if any locked
attribute changes, any generated text is substituted for the real label, or the
intended quantity, included parts, scale or use becomes unclear.

    The non-negotiable fields are the exact SKU/variant, locked attributes, permitted edit and rejection conditions. You may omit decorative details such as a named interior style. Never ask the model to infer a missing colour, reverse view, component, quantity, dimension or claim.

    A compact mobile version

    If a mobile interface makes long prompts awkward, keep product identity and the stop rule:

    Edit the attached photos of exact SKU [ID/variant]. Change only [area]. Preserve its
exact shape, proportions, colour, material, finish, pattern, label text, quantity,
included parts and verified scale. Create [scene/output/crop]. Add no product parts,
props that look included, text or claims. Reject any result that changes the product.

    Shorter is acceptable; vague is not. “Make this premium, cinematic, ultra-realistic and 8K” says almost nothing about the sale item.

    Completed editorial example

    The following demonstrates the grammar with the fictional terracotta jar used in GPTWala’s parent guide. The reference exists only as an editorial image; there is no physical sale SKU, so the output must remain an illustration and cannot become product proof.

    SOURCE TRUTH
Use the supplied front reference of fictional editorial jar EDU-JAR-01 as the only
source of visual identity. Lock its tall cylindrical silhouette, matching terracotta
lid and knob, matte warm-terracotta body, one raised horizontal band around the upper
body, no handles, one visible jar, and the exact front-facing proportions shown.

PERMITTED CHANGE
Change only the canvas outside the jar. Do not regenerate, reshape, recolour, relabel,
resize, sharpen or add texture to the jar.

SCENE AND OUTPUT
Create a horizontal editorial lifestyle illustration. Place the jar on a warm neutral
kitchen shelf, viewed at the same camera angle, with soft daylight from the left,
a small grounded contact shadow, restrained background objects and clear negative
space on the right. Use a 16:9 crop.

NEGATIVE + ACCEPTANCE CONDITIONS
No second jar, food claim, ingredient, logo, label, offer text, human hand or accessory
touching the jar. Reject any result that adds handles or changes the lid, knob, band,
body colour, finish, silhouette or perceived size. Keep the caption “fictional editorial example”.

    That example is useful for learning the syntax—not for proving preservation. For a commercial image, replace the fictional input with a photographed exact SKU and a real product truth card.

    Official controls differ by tool and version. As reviewed on 11 August 2026, OpenAI’s ChatGPT Images help says a user can upload an existing image, describe an edit and select a specific area. It also warns that highlights are not always precise and an edit may extend outside the selection. Google Product Studio’s help describes scene generation around a product image and warns that experimental features can produce unexpected output. A protected area is a useful control, not a warranty.

    Before you copy a prompt, make a product truth card

    The prompt should be assembled from a record, not from memory. Photograph the exact item from enough angles, then let the product owner or SKU expert complete this card.

    Product truth field Verified value Reference file or physical check Stop-ship if changed?
    SKU and variant [enter] [filename / item in hand] Yes
    Silhouette and proportions [enter] [front + side] Yes
    Colour and colourway [enter] [controlled reference] Yes
    Material and finish [enter] [macro/detail] Yes
    Pattern, weave or surface [enter] [detail] Yes
    Label, logo and readable text [enter] [label close-up] Yes
    Quantity sold [enter] [offer record] Yes
    Included parts/accessories [enter] [complete pack shot] Yes
    Dimensions and scale cue [enter] [measured record] Yes
    Permitted edit [enter] [approved brief] —
    Intended image role/channel [enter] [approved brief] —
    Named reviewer [enter] [approval log] —

    Universal locked attributes

    Lock the SKU, variant, silhouette, proportions, colour, material, finish, pattern, label/logo, quantity, components and scale whenever they affect what the buyer receives. Also lock any small feature that distinguishes one variant from another: cap type, handle shape, port location, fastening, seam, edge profile or pack size.

    Generated label text is untrusted even when it looks readable. Keep the photographed label layer whenever possible; otherwise add text manually from an approved source and review it at full size.

    Category-specific locked attributes

    Category Add these locks Do not infer
    Jewellery Stone count, setting, prongs, metal tone, clasp, chain length and proportions Hallmark, purity, weight, stone identity or size
    Apparel Weave, print, motif/border placement, embroidery, stitching, cut, colour, drape and supplied size Exact fit on a body, unsupplied back view or colourway
    Packaged goods/cosmetics Pack shape, cap/pump, closure, label, net quantity and approved claims Ingredients, benefits, certification or revised artwork
    Footwear Upper, sole pattern, stitching, eyelets, fasteners and colourway Comfort, grip, fit, material performance or unseen outsole
    Manufactured component Holes, ports, threads, fasteners, dimensions, finish and included pieces Internal construction, load, capacity, compatibility or tolerance

    What a prompt cannot recover

    Stop and recapture if a label is unreadable, an edge is clipped, colour is visibly wrong, a reflective surface hides its geometry, a reverse side is missing, or dimensions are unknown. A longer prompt cannot recreate evidence that was never supplied. Build the source and approval workflow before prompting, then return here.

    Catalogue and main-image prompts

    Catalogue images answer “What exactly will I receive?” Creative freedom should be low. A prompt does not make an image compliant with Amazon, Google, Flipkart, Meesho or any other destination. Check the current rule and category in the seller account before use.

    Google’s current main product image guidance requires an actual, accurate product image, rejects generic or promotional imagery in many cases, and asks merchants to show the correct variant, colour, pattern and material. It also requires generative-AI metadata to remain embedded. Treat those as destination checks, not universal specifications for every platform.

    Prompt 1 — Clean catalogue background, preserve lane

    Status: Template—verify with your SKU and current destination rules. Use this only when the tool can keep the real product and change the area around it. For a main image, the channel’s current rules outrank the scene description.

    SOURCE TRUTH
Use the attached front and 45-degree photos of exact SKU [ID], variant [name], as the
only product source. Lock the exact outer edge, proportions, colour [verified value],
material/finish [value], pattern [value], photographed label and text, one sale unit,
all included parts [list], and verified dimensions [value].

PERMITTED CHANGE
Replace only the pixels outside the product with [pure white / destination-approved
neutral background]. Preserve the real product layer and edge. Do not redraw,
reconstruct, recolour, retouch or upscale product details.

SCENE AND OUTPUT
Create a clean catalogue image for [channel and image role], keeping the supplied
camera view. Centre the product with [approved margin/crop], even neutral light and
only a restrained physically plausible contact shadow if the destination allows it.
Output [aspect ratio and minimum size checked on publication date].

NEGATIVE + ACCEPTANCE CONDITIONS
No props, extra units, hands, text, badges, border, watermark, invented reflection or
unlisted accessory. Reject if the edge, colour, texture, label, quantity, included
parts, scale or crop changes, or if the product is partially hidden.

    Compact version: “Keep exact SKU [ID/variant] untouched. Replace only the background with [current channel-approved background]. Preserve edge, shape, colour, texture, label, one sale unit, included parts and scale. No prop, overlay, extra unit or redraw. Reject any product change.”

    If the tool redraws the label or edge while replacing the background, do not keep regenerating. Restore the real layer with a controlled mask, use a manual cutout or move to a hybrid editor.

    Prompt 2 — Transparent cutout with natural edge control

    Status: Template—verify with your SKU and a tool that supports transparent output or controlled masking. Automatic cutouts are especially risky for glass, chrome, fine chains, fur, translucent packs, wispy fabric and soft shadows.

    SOURCE TRUTH
Use the supplied high-resolution image of exact SKU [ID/variant]. Lock every visible
product pixel and boundary, including [thin edge/chain/fibre/transparent area], exact
colour, material, label, quantity, included parts and the existing product geometry.

PERMITTED CHANGE
Remove only the background outside the verified product boundary. Output transparency
outside that boundary. Preserve legitimate openings, translucent areas and fine detail;
do not invent missing edges or fill holes.

SCENE AND OUTPUT
Create a transparent PNG master for controlled downstream design, at the source
resolution and original viewpoint. Keep a separate untouched source file.

NEGATIVE + ACCEPTANCE CONDITIONS
No halo, jagged edge, clipped chain/fibre, filled opening, added reflection, softened
label, reconstructed corner or generated shadow. Reject if the mask cannot separate
the edge confidently. Route uncertain edges to manual masking or real retouching.

    A transparent file is an editing asset, not proof that its edge is correct. Inspect it at 100–200% over light, mid-tone and dark temporary backgrounds before approval.

    Lifestyle and additional-image prompts

    Lifestyle images answer “Where might this product fit?” They can use more context than catalogue images, but the product, offer and use must remain truthful. Google’s lifestyle image guidance describes real-world context as a lifestyle role, bars promotional overlays in that feed field and requires generative-AI metadata to be preserved. Other channels may classify the same asset differently.

    Prompt 3 — Neutral tabletop lifestyle scene, contextualise lane

    Status: Template—verify with your SKU. Use a restrained setting before attempting a complex room or campaign.

    SOURCE TRUTH
Use the attached exact SKU [ID/variant] product layer and reference views. Lock its
shape, proportions, [colour], [material/finish], [pattern], real label, [quantity],
[included parts] and verified dimensions [value].

PERMITTED CHANGE
Change only the background, supporting surface, surrounding light and contact shadow.
Keep the product layer, view and scale unchanged.

SCENE AND OUTPUT
Place the product on a [warm neutral wood / matte stone / plain counter] in a simple
[home/workshop/retail] setting. Use eye-level or slight 15-degree-down viewpoint,
soft daylight from [left/right], a short contact shadow matching that light, and a
[4:5 / 1:1 / 16:9] crop. Keep background depth subtle and props visually secondary.

NEGATIVE + ACCEPTANCE CONDITIONS
No duplicate product, hand, ingredient, accessory, logo, readable invented text,
badge or prop that looks included in the sale. No use or performance implication.
Reject product drift, floating contact, impossible reflection or misleading scale.

    Choose props by exclusion as much as by style. A lid, charger, serving spoon, chain extender or refill placed too close to the product may look included even if the prompt calls it “decor”.

    Prompt 4 — Scale-aware in-use context

    Status: Template—verify with your SKU. Only use this prompt after measuring the product and supplying a trustworthy scale reference. Never ask the model to “make it look compact” or “show its generous size.”

    SOURCE TRUTH
Use exact SKU [ID/variant] from the supplied product views. Its verified dimensions are
[H × W × D / diameter / length] and its verified quantity is [value]. Use the supplied
[ruler/fixture/known object/body-area] reference only as a scale cue. Lock the product’s
shape, colour, material, label and included parts.

PERMITTED CHANGE
Create context around the unchanged product. Do not resize, stretch, crop, rotate into
an unsupported view or modify the product to fit the scene.

SCENE AND OUTPUT
Show the product [placed/held/worn/installed] in the verified use context [description],
from [viewpoint], with the product dimensions remaining consistent with the supplied
scale cue. Use [lighting], believable contact/occlusion and [aspect ratio].

NEGATIVE + ACCEPTANCE CONDITIONS
No unsupported load, capacity, fit, safety, medical, food-contact, waterproof or
compatibility implication. Add no body part unless rights and suitability are cleared.
Reject if scale cannot be verified or the context changes what buyers may expect.

    If fit or safety is a material buying claim, prefer a real demonstration and measured caption. A plausible hand, room or model can make the wrong size feel convincing.

    Prompt 5 — Seasonal or regional campaign scene

    Status: Template—verify with your SKU and review cultural context. Specify one occasion and a restrained visual vocabulary. “Indian festival background” is too vague and often produces clutter or mismatched symbols.

    SOURCE TRUTH
Use attached exact SKU [ID/variant]. Lock its product layer, shape, colour, material,
surface, label, quantity, included parts and verified scale.

PERMITTED CHANGE
Change only the setting, ambient light and secondary decor around the product. Do not
alter the product to match the occasion.

SCENE AND OUTPUT
Create a [occasion/region]-appropriate campaign setting using only [two or three
specific, reviewed decor cues], [approved palette], [surface], [light direction] and
[crop]. Keep the product dominant with clean negative space for verified copy to be
added later in design software.

NEGATIVE + ACCEPTANCE CONDITIONS
No generated offer, price, discount, review, certification, gift claim, religious
symbol, person, extra product unit or brand mark unless separately approved and
supplied. Reject stereotyped, mixed or disrespectful cues and any product change.

    Add price, discount, dates and terms later from an approved offer sheet. Do not depend on an image model to spell, calculate or substantiate them.

    Prompt 6 — Consistent multi-SKU catalogue series

    Status: Template—verify each SKU separately. Consistency means reusing a scene specification—not asking the tool to invent missing variants.

    SOURCE TRUTH
This run is only for exact SKU [ID], variant [name]. Use its own supplied front,
45-degree, back and detail views. Lock its individual shape, proportions, colour,
material, finish, pattern, label, quantity, included parts and dimensions. Do not use
another SKU’s product pixels or infer a colourway.

PERMITTED CHANGE
Reuse only the approved series specification: [background], [surface], [camera view],
[crop], [light direction], [shadow style] and [margin]. Change no product attribute.

SCENE AND OUTPUT
Create one [catalogue/additional] asset matching series ID [SERIES-ID], at [aspect ratio
and size], while showing this exact SKU clearly. Name the candidate [SKU_ROLE_V01].

NEGATIVE + ACCEPTANCE CONDITIONS
No merged variants, borrowed labels, averaged proportions, extra units or accessories.
Reject any product drift or visual inconsistency that hides a distinguishing feature.
Approve this SKU independently before starting the next SKU.

    Store source, prompt, output, rejection reason and approval per SKU. A beautiful batch is still unusable if one jar has the wrong cap or one kurta has an invented border.

    Specialist product prompts

    These templates narrow the job for higher-risk categories. They do not replace real proof images, category expertise or destination checks.

    Prompt 7 — B2B manufacturer or dealer detail image

    Status: Template—verify with engineering or product records. Use it to reveal a photographed detail, not to fabricate an internal cutaway or performance demonstration.

    SOURCE TRUTH
Use exact manufactured SKU [part/model ID] and the supplied overall, side and macro
detail photos. Lock external geometry, hole/port/thread/fastener count and positions,
finish, colour, visible markings, verified dimensions, one sale quantity and included
components [list].

PERMITTED CHANGE
Change only the background, crop and non-product annotation space. Preserve the real
product and supplied macro detail. Do not invent an unseen interior or mating part.

SCENE AND OUTPUT
Create a dealer-catalogue detail image that keeps [verified feature] clearly visible
from the supplied viewpoint, on a neutral technical surface, with even light, truthful
scale and empty space for a manually added dimension callout. Output [aspect ratio].

NEGATIVE + ACCEPTANCE CONDITIONS
No cutaway, load, flow, capacity, compatibility, tolerance or durability claim. No
added bolt, port, tool or assembly part. Reject any changed geometry, count, marking,
dimension cue or implied included component.

    Dimensions and arrows should be added manually from the approved technical record. Do not let the image generator create numerals or engineering labels.

    Prompt 8 — Apparel secondary image with model or context

    Status: Template—verify with your exact garment, model rights and tool terms. Use the result as additional or lifestyle context, not as proof of exact fit, fall or drape.

    SOURCE TRUTH
Use the supplied front, back, flat-lay and macro photos of exact apparel SKU [ID],
colourway [name] and size [size]. Lock base colour, fabric appearance, weave, print,
motif sequence, border width and placement, embroidery, neckline, sleeve, hem,
stitching, closures and all included pieces.

PERMITTED CHANGE
Add only the approved model/setting around the garment using a workflow whose rights
and consent terms have been reviewed. Do not redesign, tailor, lengthen, shorten,
smooth away, recolour or invent an unsupplied garment view.

SCENE AND OUTPUT
Create a secondary lifestyle image with [approved model description and pose],
[setting], [camera view], [lighting] and [crop]. Keep the complete garment visible and
provide a separate crop for border/embroidery detail if needed.

NEGATIVE + ACCEPTANCE CONDITIONS
No invented print, border, sleeve, blouse piece, pocket, lining, accessory, body-shape
claim or extra colourway. Do not describe the result as exact fit proof. Reject if
motif, construction, colour, coverage, fall or included pieces differ from the SKU.

    Keep real flat-lay, reverse and detail images beside any model image. A model scene can communicate styling; it cannot establish the exact experience of every body or size.

    Prompt 9 — Jewellery contextual image

    Status: Template—verify against the item in hand and real macro photographs. Fine geometry and reflections make jewellery one of the easiest categories to alter invisibly.

    SOURCE TRUTH
Use the exact jewellery SKU [ID/variant] product layer plus supplied front, reverse,
clasp and macro references. Lock item count, stone count and arrangement, setting and
prongs, metal tone, surface finish, chain/bracelet length from the verified record,
clasp type, pendant/earring proportions and every visible construction detail.

PERMITTED CHANGE
Create only the background, supporting surface, restrained reflection and surrounding
context. Preserve the photographed jewellery layer. Do not redraw stones, chain links,
settings, hallmark or clasp.

SCENE AND OUTPUT
Place the product in a minimal [velvet/stone/plain skin-safe approved] context with
soft controlled light, the supplied camera view, truthful scale and [aspect ratio].
Keep the jewellery unobstructed and include a separate real macro proof image.

NEGATIVE + ACCEPTANCE CONDITIONS
No added sparkle that hides detail, extra stone, changed setting, thickened chain,
different metal colour, invented hallmark, purity/weight claim or misleading body
scale. Reject any uncertain count, geometry, reflection, mark or proportion.

    Never infer purity, weight, hallmark or stone identity from appearance. Those facts belong in verified product data, not in a generated visual.

    Ad-creative prompt

    An ad image can use a more deliberate crop and negative space, but it still cannot invent the product, offer or evidence.

    Prompt 10 — Ad-creative crop from an approved product master

    Status: Template—verify with your SKU, approved master and actual ad placement. Use design software to add verified copy after the image is approved.

    SOURCE TRUTH
Use approved product master [asset ID] for exact SKU [ID/variant]. Lock all product
pixels, shape, colour, material, label, quantity, included parts and scale. The master,
not a prior generated ad, is authoritative.

PERMITTED CHANGE
Extend or replace only the canvas outside the approved product. Reposition the intact
product layer within the crop if needed; do not generate a new product angle.

SCENE AND OUTPUT
Create a [Meta/website/WhatsApp] creative background for [audience/use case], with
[surface/context], [light], strong product visibility and clean negative space on
[side] for manually added approved copy. Output [placement aspect ratio and safe area].

NEGATIVE + ACCEPTANCE CONDITIONS
No generated price, discount, star rating, testimonial, badge, before/after proof,
guarantee, scarcity, certification, benefit claim or extra item. Reject any product
change, false use implication, confusing quantity or insufficient safe space.

    The Indian government’s Consumer Protection Act FAQ explains that a misleading advertisement can falsely describe a product or mislead consumers about its nature, substance, quantity or quality. This is not legal advice; it is a practical reason to keep product and offer truth inside the creative workflow.

    Product layer locked while background, crop, lighting treatment and approved props remain inside the permitted edit boundary

    Original GPTWala edit-boundary diagram. “Locked” describes the instruction and workflow control—not a guarantee. Compare the output with the real product before approval.

    A prompt repair ladder when the product changes

    Do not add more adjectives to a failing prompt. Reduce uncertainty and strengthen control. Move down this ladder once per failed review.

    Step Action Example Stop condition
    1 Name the exact defect “The output changed the six holes to five; preserve all six in their supplied positions.” If another identity field changes
    2 Reduce permitted change Replace a complex room with a plain surface and one light direction If the product is still redrawn
    3 Protect the product layer Select/mask only the background; composite the approved real product layer If the control bleeds into the product
    4 Supply missing evidence Add side, back, macro, label or scale reference If the evidence is still incomplete
    5 Change workflow/tool Move from full-frame generation to local edit, layer compositing or manual retouching If fidelity remains inconsistent
    6 Stop AI generation Use real photography or a hybrid asset Immediately for unresolved stop-ship truth

    Repeatedly writing “do not change the product” is not a substitute for a better source, a narrower edit or a protected layer. A product-truth audit should follow every attempt.

    How to test a prompt before using it across your catalogue

    Run a five-stage prompt ladder on one owned SKU before you batch anything. Keep the source, tool, model/version, date, output settings and attempt budget fixed. This article does not publish fabricated results: as of 11 August 2026, the templates above remain marked “Template—verify with your SKU.”

    Stage Prompt/control change What to record
    1 Vague baseline: “Make this product photo premium on a lifestyle background.” Every identity, offer, geometry, text, material and scale defect
    2 Add exact SKU and locked attributes Which defects disappear, persist or newly appear
    3 Add permitted-change boundary and rejection conditions Whether product pixels still drift
    4 Add local selection/mask or protected real product layer, if supported Selection bleed, edge defects and edit-control limits
    5 Human QA and stop decision Approved, repair, recapture, change workflow or stop

    Use the same attempt cap at each stage—such as three candidates—to avoid giving the preferred method unlimited retries. Do not select only the prettiest output. Record all candidate outcomes and calculate:

    • Product-truth pass rate: outputs with zero stop-ship product/offer errors ÷ all outputs reviewed.
    • First-pass approval rate: outputs approved without repair ÷ all outputs reviewed.
    • Rework minutes per approved asset: total correction time ÷ approved assets.
    • Cost per approved asset: tool, operator, review and rework cost ÷ approved assets.

    The best prompt is the one that contributes to repeatable approved output for your SKU. It may not be the longest or most visually dramatic prompt.

    Three illustrative Indian business adaptations

    These are fictional training records—not merchant case studies, tool tests or outcome claims. Replace the values only after checking the real product and source pack.

    Rajkot manufacturer: dealer detail image

    Training truth card: EDU-COUPLING-50; stainless-steel coupling; 50 mm verified outer diameter; six equally spaced visible bolt holes; one coupling; no bolts included.

    Use the supplied front, side and macro references of exact training SKU EDU-COUPLING-50. Change only the background to neutral charcoal and preserve the 50 mm scale cue, cylindrical geometry, stainless finish, six hole positions, one-unit quantity and visible marking. Create a 4:5 dealer detail image with even side light and blank space for a manually added dimension line. Add no bolt, mating part, cutaway, capacity or compatibility claim. Reject any change to hole count, geometry, marking, scale or included parts.

    Surat apparel wholesaler: secondary kurta image

    Training truth card: EDU-KURTA-INDIGO-M; indigo cotton kurta; white repeated motif; 35 mm verified hem border; three-quarter sleeve; one kurta; no dupatta included.

    Use the supplied front, back and motif close-ups of exact training SKU EDU-KURTA-INDIGO-M. Add only a rights-cleared standing model and plain limewash-wall context. Preserve the indigo colour, white motif sequence, 35 mm hem border, neckline, three-quarter sleeves, stitching and one-piece offer. Use soft daylight and a full-garment 4:5 crop. Add no dupatta, jewellery, pocket, alternate print or fit claim. Reject changed colour, motif, border, cut, drape or implied included item. Keep real flat-lay/detail images as proof.

    Local packaged-product retailer: festive additional image

    Training truth card: EDU-SPICE-TIN-100; one 100 g round spice tin; matte ochre body; black lid; photographed label retained; no gift box included.

    Use the approved real product layer for exact training SKU EDU-SPICE-TIN-100. Change only the setting to a restrained Diwali tabletop with one warm brass lamp in the distant background and a few marigold petals outside the product boundary. Preserve the round ochre tin, black lid, real label, 100 g net quantity, one-unit offer and scale. Leave clean space for approved copy to be added later. No generated discount, ingredient, certification, gift box, extra tin or altered label. Reject product drift, confusing quantity or decor that implies inclusion.

    Notice how the adaptations change the task and risk—not just the industry noun. The manufacturer needs verified geometry; the apparel seller needs construction and offer clarity; the retailer needs pack and promotion truth.

    What prompts cannot solve

    A prompt cannot solve:

    • poor, clipped or colour-inaccurate source photography;
    • missing reverse, label, macro or scale evidence;
    • a tool that redraws the product despite local instructions;
    • exact colour calibration across capture, monitor and buyer screen;
    • model, location, trademark or uploaded-design rights;
    • privacy and retention questions for confidential catalogues;
    • current marketplace, category or regulated-product restrictions;
    • unsupported product performance, fit, safety or health claims;
    • final approval by someone who knows the exact sale item.

    If the source is weak, return to the phone-to-approved workflow. If the output looks right but you cannot verify it, run the full product-accuracy audit. If the destination rule is unclear, check the current seller documentation instead of adding “marketplace-ready” to the prompt.

    Turn the prompt into a repeatable content system

    A prompt pack is useful only when it sits inside a process: verified product → approved image → channel-ready content → distribution → enquiry follow-up. In GPTWala’s DAA framework, prompt-led assets support the AI Content Creation layer; they still need a Digital Presence and a practical WhatsApp advertising and follow-up system.

    Join the GPTWala workshop to learn how these pieces connect, including the taught ₹100/day WhatsApp ads setup. ₹100/day is a starting-budget concept taught in the workshop—not a promise of leads, sales or profitability.

    Frequently asked questions

    What is the best prompt for AI product photography?

    The best starting prompt identifies the exact SKU, lists locked attributes, limits the permitted edit, specifies the image job and defines rejection conditions. It is only “best” after it produces repeatable approved assets for your own SKU under a dated test. No universal wording guarantees product preservation.

    Do negative prompts stop an AI tool from changing the product?

    No. Negative constraints reduce ambiguity, but the tool may still alter product pixels, especially during a full-frame generation or an imprecise selection edit. Use the real reference, narrow the editable area, compare side by side and reject material drift.

    Can I use the same prompt in ChatGPT, Product Studio and other image tools?

    Reuse the same semantic brief—source truth, permitted change, scene and rejection conditions—but adapt it to the controls the current tool actually supports. Do not copy invented parameters between tools. Recheck official help after model or editor updates.

    Does ChatGPT support editing a product reference image?

    As reviewed on 11 August 2026, OpenAI’s official help says ChatGPT Images can edit an uploaded image, accept a described change, target a selected area and use a chosen aspect ratio. It also says selections may not be precise and edits can extend beyond the highlighted area. Verify the exact interface, plan and terms when you use it.

    Why does AI keep changing labels and packaging text?

    Image models may regenerate visual text rather than preserve the photographed label. Treat every generated character as untrusted. Retain the real label layer or add verified text manually from the approved artwork, then review at full resolution.

    Can these prompts make an Amazon, Flipkart or Google main image?

    They can direct a candidate edit; they cannot certify compliance. Use the real sale item, then check the current channel, country and category rules in the seller surface. Main-image, additional-image and lifestyle roles are not interchangeable.

    Are AI model images safe for apparel and jewellery?

    They are higher-risk secondary assets. Apparel can drift in motif, border, construction, colour, fit and drape; jewellery can drift in stone count, setting, clasp, metal tone and scale. Keep real main and detail proof, and route these categories through specialist review.

    Is one product photo enough for these prompts?

    Usually not for commercial approval. A single front image cannot prove the back, side, label, clasp, ports, included pieces or dimensions. Capture the views required to verify the product before generation.

    When is a prompt ready for batch production?

    Only after a controlled pilot records the tool/model/date, fixed attempt count, truth defects, approvals, rework time and cost per approved asset. One attractive output is not a repeatable system.

    Sources and review method

    This prompt pack was researched and reviewed on 11 August 2026. Tool interfaces, model behaviour, terms and marketplace image rules can change. Recheck any named tool within 24 hours of publication and after major product updates; recheck destination rules at least every 90 days and on the day of final upload.