The product photography problem isn’t a shortage of AI tools. It’s that studio product shot, white background gets you a generic floating blob with no shadow logic and no material believability, so it reads fake the second you drop it into a real listing. This tutorial gives you the five-slot prompt scaffold real product photographers think in (subject, material, finish, background, lighting), the camera-language vocabulary that signals “shot on glass” instead of “rendered in software”, the iterate-one-variable rule that stops you burning credits on chaotic regenerations, and a clear answer to which June-2026 model to reach for per product type.
TL;DR
- Use a 5-slot prompt: subject + material + finish + background + lighting. Vague prompts produce vague, fake-looking results.
- Add camera language (lens, aperture, framing). This is the single biggest signal that separates “render” from “photograph”.
- Iterate one variable per pass. Change lens, lighting, and background at once and you’ll never know what helped.
- Model pick (as of June 2026): Google Nano Banana Pro (Gemini 3 Pro Image) for photoreal product shots and readable labels; GPT Image 2 in ChatGPT when packaging copy or on-pack typography must be legible; Seedream 5.0 for cheap, high-volume stylized lifestyle scenes.
- Reflective / glass / chrome products are the hard category. Always feed a real reference photo (image-to-image) and write explicit reflection prompts.
- Never use AI for the Amazon main image. Amazon policy requires the real product on a pure-white background (RGB 255,255,255) filling 85%+ of the frame.
Which model to use (June 2026)
The tool you pick matters more than any prompt trick. As of June 2026 these are the three engines worth your time, plus two dedicated e-commerce apps that wrap them with batch tooling.
| Tool / model | Where it lives | Best for | Price (as of June 2026) |
|---|---|---|---|
| Nano Banana Pro (Gemini 3 Pro Image) | Gemini app, Google AI Studio, API | Photoreal product shots, lighting/reflection realism, readable labels; native 1K with upscale to 2K/4K | Google AI Pro $19.99/mo; API ~$0.043/image |
| GPT Image 2 | ChatGPT (Plus and up) | On-pack typography and packaging copy that must be legible (~99% character accuracy), product mockups | ChatGPT Plus $20/mo |
| Seedream 5.0 | API and third-party apps | Cheap, high-volume stylized lifestyle scenes, multi-image editing | API ~$0.018–0.05/image |
| Photoroom | Web + mobile app | Background removal, instant scene staging, batch e-commerce export | Free (1 HD export/day) / Pro $9.99/mo / Business $29.99/mo |
| Pebblely | Web app | Lifestyle backgrounds from one product photo, no prompt skill needed | 40 free/mo / Lite $9 (30 img) / Basic $19 (200) / Pro $39 (500) |
Rule of thumb: if your product has text on the package (a label, a back-of-pack ingredient list, a logo that must stay sharp), start in GPT Image 2 — it’s the only model that reliably keeps typography legible. For everything else where you want it to look genuinely photographed, Nano Banana Pro is the strongest default in June 2026 because its “Thinking” pass adjusts lighting and reflections to match each edit instead of fighting you. If you’re producing hundreds of stylized scenes on a budget, Seedream 5.0 through an API is the cheapest production-grade option.
Who this is for
E-commerce sellers building listings, indie product founders shipping launches solo, Etsy / Shopify owners doing weekly fresh creative, and content teams generating placeholder imagery while real photography is scheduled. It’s especially worth it for variants — same product, six color options — where reshooting each is overkill.
When this is NOT the right tool
Use real photography or photogrammetry for:
- The Amazon / Etsy main “hero” image. Amazon’s automated checks require the actual product, a pure-white RGB 255,255,255 background, the product filling 85%+ of the frame, and no added text, logos, or watermarks. A generated image will get the listing suppressed. AI is fine for secondary lifestyle and detail slots.
- Regulated categories where ad imagery is inspected for misleading depiction: supplements, medical devices, financial products.
Before you start
- Pull 1-2 reference photos of your real product. Image-to-image always beats text-only for fidelity, and on reflective products it’s the difference between usable and unusable.
- Decide finished platform sizes first: Amazon needs 2000px+ square (1000px minimum for zoom), Instagram 1080×1350, Shopify product 2048×2048. Set generation size accordingly — Nano Banana Pro renders 1K natively and upscales to 2K/4K, so generate large and downscale, never upscale a small output.
- Pick a brand visual mode in one sentence: “warm-lit kitchen lifestyle”, “cool studio minimal”, “moody product-noir”. This becomes the anchor for every subsequent prompt.
- Have brand hex codes ready. “In brand colors” without hex drifts every time.
The 5-slot prompt, step by step
- Fill all five slots: subject + material + finish + background + lighting. Example:
ceramic mug, glazed stoneware, matte finish with subtle texture, seamless paper backdrop in warm cream, soft key light from camera-left, gentle fill from a reflector camera-right. - Add camera language: lens + aperture + framing. Example:
35mm, f/2.8, shallow depth of field, three-quarter angle, eye-level. Camera vocabulary is the single biggest signal that separates a render from a photograph. - Specify shadow direction and softness. Example:
soft shadow falling to lower-right at 30 degrees, edge softness about 4 inches. Vague shadow language produces vague shadows. - Iterate one variable per pass. Don’t simultaneously change lens, lighting, and background — you’ll never know what helped. Pick one, generate 4, evaluate.
- For reflective / glass / chrome, add explicit reflection prompts. Example:
subtle environment reflection on top surface, no hot-spot blowout, soft gradient highlight on side. Reflective surfaces are AI’s weak spot; combine this with an image-to-image reference photo. - Generate 6-8 variants per prompt. The first output is rarely the keeper. Mark each “usable / needs retouch / reject” so you build vocabulary about what your prompt actually does.
- Retouch lightly in Photoshop / Affinity / Photopea. Clean stray pixels, level brightness, sharpen the subject. AI does about 90% of the work; that last 10% is what separates “AI shot” from “studio shot”.
For the field-by-field prompt scaffold (key light, fill, lens, ratio, brand mood), see our writing a product image prompt walkthrough before your first run.
First-run exercise
- Pick one real product you actually need to photograph this week. Real listings surface real decisions that practice runs miss.
- Run the 5-slot scaffold once end-to-end in Nano Banana Pro (or GPT Image 2 if the product has on-pack text). Save the output even if rough.
- Place the image into the actual listing layout — Amazon secondary thumbnail, Shopify product page. Most AI images look fine isolated and reveal problems in context.
- Second pass: change ONE variable. The most common improvement is the shadow specification or the aperture (depth of field).
Quality check
- Shadow physics: direction, softness, and color should match the stated light source. Floating products read as fake immediately.
- Material believability: ceramic should show slight surface variation, wood should show grain, metal should show micro-scratches and reflection.
- Edge artifacts at 100% zoom: halos and AI fringe show up around fine details — handles, lids, transparent areas.
- One believable light setup: two contradicting sources (sun from the left, hard shadow on the right) is the most common giveaway.
- Color accuracy: compare to the real product with a color picker. If brand colors don’t match, your hex prompt was ignored — nudge brightness/saturation in post.
How to reuse this workflow
- Save winning prompts by product category. All mugs, all candles, all bottles get the same template; only the product description changes.
- Build a small library of reject reasons: drifted logo, melted handle, contradictory shadow, plastic-looking ceramic. Naming failures speeds the next prompt.
- Pair every saved prompt with a real reference photo so the next person (or future you) can image-to-image without losing fidelity.
- Re-test the template every 4-6 weeks. Image models ship fast — GPT Image 2 landed in April 2026 — and the workaround you needed last month may now be the default.
Common mistakes
- Vague “product on white” produces a generic floating product with no shadow logic. Always specify lighting direction and shadow geometry.
- Generic “studio prompt” produces the same default-lit AI ceramic regardless of your actual product. Specify material and finish.
- Skipping camera language makes results read as renders, not photographs.
- Iterating multiple variables means you never learn which fix worked. One variable per pass.
- Judging from a single output. Generate at least 6 before evaluating.
- Using the wrong model for the job. Photoreal label-heavy product in a model that mangles text, or paying for a stylized engine when you needed a clean studio shot.
- Skipping the retouch step. AI does 90%; the final 10% Photoshop pass is what separates “AI shot” from “studio shot”.
FAQ
- Can I use these for Amazon main images? No. Amazon policy requires the actual product on a pure-white RGB 255,255,255 background, filling 85%+ of the frame, with no added text or logos; generated mains get suppressed. Use AI for secondary lifestyle and detail images only.
- Which model is best for product photos right now? As of June 2026, Nano Banana Pro (Gemini 3 Pro Image) is the strongest default for photoreal shots and lighting realism. Switch to GPT Image 2 when the package has text or a logo that must stay legible — it hits roughly 99% character accuracy. Use Seedream 5.0 for cheap, high-volume stylized scenes.
- What about reflective / glass products? The hardest category — expect 3-5x more rejects. Always start from a real reference photo via image-to-image and write explicit reflection prompts.
- Can I match my exact brand colors? Include hex codes in the prompt and verify the output with a color picker. Drift is common; nudge brightness/saturation in post.
- How long should one product take? First time end-to-end: 60-90 minutes. With saved templates: 15-20 minutes per SKU.
- Cheapest way to do this at volume? A dedicated app like Photoroom (Pro $9.99/mo) or Pebblely (40 free images/month) handles batch staging without prompt skill, or call Seedream 5.0 directly at roughly $0.02-0.05 per image.