An image generation API workflow for ecommerce catalogs

Match each catalog job to the right model

You can produce catalog-ready product images programmatically with two live models behind one prepaid apiToken.sale key: Nano Banana 2 (gemini-3.1-flash-image) on the native Gemini route and GPT Image 2 (gpt-image-2) on the OpenAI Images route. The right choice is per job, not per store: Nano Banana 2 gives explicit 1K/2K/4K sizes, broad aspect-ratio control and up to 14 reference inputs, while GPT Image 2 gives an OpenAI Images edit flow with one to five PNG, JPEG or WebP references for teams already on that client.

GPT Image 2 publishes transparent cutouts when you send background=transparent: the PNG has a real alpha channel. size is a proportion preset (auto/1:1 ~1254×1254, 3:2, 2:3, 4:3, 16:9, 9:16), not a 1024-pixel or 4K lock. If a fixed pixel size is a hard storefront requirement, reject that route at design time instead of repairing every output later.

Editing route contracts: references, formats, output shapes →

See also: Edit images through GPT Image 2 or Nano Banana 2

Two wire protocols behind one prepaid key

The same key works for both routes, but the request shapes are not interchangeable. Nano Banana 2 is called with the Gemini generateContent shape and the x-goog-api-key header; the response carries the picture as a base64 inlineData image part. GPT Image 2 uses the OpenAI Images endpoints with Authorization: Bearer and returns one non-streaming base64 PNG. Size and quality controls differ too: imageConfig.imageSize (1K/2K/4K) plus aspectRatio on the Gemini side, the published background/quality/size values opaque-or-transparent/low/proportion-preset on the OpenAI side.

# Nano Banana 2 — native Gemini route, explicit size and ratio
curl https://router.apitoken.sale/v1beta/models/gemini-3.1-flash-image:generateContent \
  -H "x-goog-api-key: $APITOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Place SKU-1042 on a warm kitchen counter, soft daylight"}]}],"generationConfig":{"responseModalities":["TEXT","IMAGE"],"imageConfig":{"imageSize":"1K","aspectRatio":"4:5"}}}'

# GPT Image 2 — OpenAI Images route, published controls only
curl https://router.apitoken.sale/v1/images/generations \
  -H "Authorization: Bearer $APITOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-image-2","prompt":"Studio packshot of a matte black bottle, seamless grey background","background":"opaque","quality":"low","size":"auto"}'

Getting started is cheap by design: new B2C accounts created with Google or GitHub receive $5 of platform bonus credit (valid on Claude, GPT and Gemini models; email-and-password accounts are not eligible), and balance top-ups accept bank card or cryptocurrency. For a catalog worker, issue a dedicated key with a lifetime spending limit and an expiration date so a runaway batch cannot drain the whole account.

A production workflow that protects the product

1. 01Normalize source photography and assign an immutable SKU/asset version to every input file.
2. 02Write the brief as allowed changes plus protected traits: silhouette, color, logo, labels and included accessories.
3. 03Generate one bounded candidate, then validate product identity, rendered text, artifacts, composition and policy compliance.
4. 04Store the original, output, prompt version, model, request ID, terminal usage and reviewer verdict as one durable record.
5. 05Publish only an optimized web derivative; never serve the API's base64 payload directly as a storefront asset.

PRODUCT: SKU-1042, matte black bottle, silver cap, logo unchanged
ALLOWED: replace the background with a warm kitchen scene
PROTECTED: silhouette, cap shape, logo spelling, label colors
OUTPUT: one centered catalog hero image; no added text or accessories

A prompt is an instruction, not a guarantee. Validation against the protected traits is what keeps a regenerated background from quietly becoming a different product.

Cost math for a real catalog run

Nano Banana 2 has a fixed image-output leg per size: 1K bills 1,120 image tokens, which is $0.0672 official and $0.0336 for a regular B2C account after the 50% discount; 2K is $0.0504 and 4K is $0.0756 on the same terms. A seasonal refresh of 500 SKUs with two 1K candidates each produces 1,000 images, so the image-output legs settle at $33.60 for regular B2C, plus bounded prompt and reference input. If validation accepts 700 of those candidates, the image-leg cost per accepted asset is $33.60 ÷ 700 ≈ $0.048 — that is the number to compare against a reshoot budget, not the raw request count.

GPT Image 2 has no honest per-picture list price: billing combines terminal text-input, image-input, cached-input and image-output usage. A worked example with round numbers — if terminal usage reports 500 fresh text tokens and 4,000 image-output tokens, the official total is 500 × $5/M + 4,000 × $30/M = $0.1225, and a regular B2C account pays exactly half, $0.06125. References add image-input at $8/M official ($4/M after the discount), so an edit brief should send only the files the scene actually needs.

For regular B2C, the 50% discount applies to the exact official generation or edit usage. Conversion, CDN storage, human review and rejected assets remain your application costs and belong in the ecommerce business case.

Budget-safe batch pipeline for whole-catalog runs →

Ship images that help storefront SEO

• Track settled cost per accepted SKU image, not request count; retries and rejects are part of acquisition cost.
• Start at the smallest delivery resolution (1K) and promote only hero assets that fail a delivery-detail check.
• Use descriptive filenames and truthful alt text describing the actual product and scene; never stuff model names or keyword lists into alt text.
• Encode responsive WebP/AVIF derivatives, keep a lossless master, and publish explicit width/height to avoid layout shift.
• Keep generated-image provenance and review evidence out of public metadata when it contains internal IDs or prompts.

Frequently asked questions

Related guides