---
name: "ai_product_photography"
description: "AI product photography for listings and ads: write the brief, capture a true reference photo on a phone budget, generate or enhance scenes without changing the product, select shots with clear rules, fix hands, labels, and reflections, and verify the final image against the real product before publishing. Use when making hero, lifestyle, detail, or ad images for Amazon, Shopify, eBay, Etsy, or social ads."
---

# AI Product Photography

## Purpose

Produce listing and ad images that sell the real product and still look like the real product when the box arrives. The core rule is simple: AI may change the scene around a product, it may not invent the product.

This skill is vendor neutral. It works with any image generator, background remover, or editor, and with no paid tool at all if the seller starts from good phone photos. It covers the full job: brief, reference capture, generation, selection, repair, and a final accuracy check that can fail an image.

## Workflow

1. Write the brief before generating anything.
   - Record: exact product name, variant, colour and finish in words, dimensions, what is included in the box, what is not included, and the one claim the image is allowed to make.
   - Record the destination first, because the rules differ. See `references/platform-specs.md` for Amazon, Shopify, eBay, and Etsy specs. A main image that passes on Shopify can be suppressed on Amazon.
   - List P0 facts that must stay exact: brand mark, product name, quantity or volume, dosage or percentage, certifications, ports, buttons, seams, and included accessories. Any P0 error is an automatic reject, covered in step 6.
   - Decide the shot set up front. Default set is six: hero on clean background, 45 degree angle, back or underside, detail close up, scale shot with a hand or a common object, and one in use shot. Full list in `references/shot-list.md`.

2. Capture a true reference photo, even if every final image will be generated.
   - Use a phone on a tripod or a stack of books. Turn off beauty filters and portrait blur.
   - Light: one large window to the side with a thin white sheet or shower curtain over it to soften the light, plus a white foam board or poster board on the shadow side to bounce light back. For small items only, a cheap light tent helps. For larger items, use a white paper sweep taped to a wall and curving onto a table, with no hard crease line.
   - Shoot the product straight on, at 45 degrees, and close on the label. Keep the label large in frame. Small label text in a wide shot is where generators fail most.
   - Shoot one frame with a grey card or a plain sheet of white paper beside the product for colour reference. Keep this frame, it is your colour truth in step 6.
   - Clean the product, remove dust and fingerprints, and photograph any real flaw on used items. Do not plan to hide flaws later.

3. Choose the route per shot, and write it in the brief.
   - Route A, real photo plus cleanup: best for hero images, used items, one of a kind items, and anything where condition matters. AI work is limited to background removal, background replacement, exposure, and dust cleanup that does not touch the product shape.
   - Route B, real product in a generated scene: mask and protect the product pixels from the reference photo, generate only the room, surface, props, and light around it, then composite. This is the default for lifestyle shots.
   - Route C, fully generated product: only for concepts, pre production mockups, and ads that are clearly not the listing hero, and only when the seller will still verify against a physical sample before use. Never use Route C as the sole image for a used item, and never on Etsy as the first image for a finished item.
   - When in doubt, pick the route with less invention. A plain true photo beats a polished false one.

4. Generate and edit with the product protected.
   - In the prompt, name the colour and finish in plain words, the material, the count of parts, and what must not appear. Example prompt frame: "Keep the product exactly as in the reference photo. Matte navy blue bottle, black cap, no extra bottles. Place it on a light oak table beside a folded white towel. Soft window light from the left. No text, no logos, no hands."
   - Ask for no text in the image. Add headlines, prices, and callouts later in a layout or design tool with real fonts. Generated text garbles, and worse, it can change a number into a different, plausible number.
   - Edit in small masked steps, one change per pass: background first, then light, then props. Compare against the reference after each pass. If a pass changes the product, undo that pass instead of prompting harder.
   - Keep hands out unless the shot needs scale or use. If a hand is needed, prefer the seller's real hand photographed holding the product over a generated hand. Generated hands fail on finger count, fused fingers, and grips that do not close on the handle.
   - Watch reflections on glass, chrome, and glossy plastic. The reflection must match the scene in the shot. A window reflection in a dark studio scene, a second product reflected that is not in frame, or a reflection with warped label text is a reject or a repaint of that region only.

5. Select shots with fixed rules, not taste alone.
   - Generate or shoot at least four candidates per slot, then score each one with `references/accuracy-check.md`. Keep the highest scoring candidate that has zero P0 errors. A pretty image with one P0 error loses to a plain image with none.
   - Check at thumbnail size first. If the product is not clear at thumbnail size, it will not work in search results, no matter how good it looks full size.
   - Check consistency across the set: same colour, same finish, same proportions, same included parts in every shot. Colour drift between shots in one listing is a common generator fault.
   - Keep one neutral, true colour shot in every set, even if the lifestyle shots are warm or graded. Buyers need one image they can trust for colour.

6. Run the accuracy check against the real product. This step can fail the whole set.
   - Put the final image and the physical product, or the reference photo, side by side. Work through `references/accuracy-check.md` line by line at 200 percent zoom.
   - Read every word and number on the label out loud and compare character by character. Check volume, weight, percentages, and product name. This is the single most common failure in flagged AI product photos.
   - Count parts, buttons, pockets, cables, and accessories. Check shape, seams, logo placement, and finish.
   - Run the bundled checker on the export files before upload: `python3 scripts/audit_images.py <folder> --platform amazon` with platform amazon, shopify, ebay, or etsy. It checks dimensions, file size, format, and, for Amazon main images, whether the corners read as pure white.
   - If any P0 item fails, fix it by compositing the original label or part back in, or reshoot. Do not ship and hope nobody zooms.

7. Export, name, and disclose correctly.
   - Export in sRGB, at the size in `references/platform-specs.md`, with a plain file name such as `navy-bottle-500ml-front.jpg`. Keep the unedited masters and the reference photos in a separate folder.
   - Write alt text that describes the real product and view, for example "Matte navy water bottle, front view, black cap".
   - Follow the destination rule on disclosure. Etsy asks sellers to state AI use in the listing description where its policy requires it. For ads, do not use an image that gives a false net impression of size, finish, results, or what is included, and do not expect a small disclosure line to fix a misleading picture. That is the FTC test: would a reasonable buyer, looking at the image, believe something material that is not true.
   - For used items on eBay, keep at least one unedited photo of the actual item showing condition and flaws. Do not use stock or generated images for used, damaged, or defective items.

## Output Contract

- A brief with destination, shot list, route per shot, and the P0 facts list.
- A final image set in the order: hero, angle, back, detail, scale, in use, each exported to the destination spec.
- A completed accuracy checklist per hero and detail image, with pass or fail on label text, numbers, colour, shape, parts count, and reflections.
- The audit script output for the export folder, with any size, format, or background warnings resolved or explained.
- A short note stating where AI was used (background only, scene, cleanup, or full generation) so the seller can make any listing disclosure the platform requires.

## Operating Rules

- Never invent the product. No added features, no removed flaws on condition shots, no changed colour, capacity, or included parts.
- Protect the label. Mask it, composite the original back in, or add text later in a design tool. Never publish a plausible substitute for factual packaging text.
- No generated testimonials, no before and after results the product cannot produce, and no scale tricks that make the product look larger or smaller than it is.
- Hands, text, and reflections get a zoom check every time, not a glance. Those three fail most often.
- One change per edit pass, compare after each pass, and keep the reference photo open the whole time.
- If the physical product is not available to check against, say so plainly and limit the work to Route A cleanup of real photos. Do not present a generated image as verified.
- Re check platform specs on the platform help pages before a large batch, because marketplaces change image rules. Specs in the references were verified at publication, not guaranteed forever.