AI product photography with an API

Why editing beats generating

If you prompt a text-to-image model for "our ceramic mug", you get a ceramic mug — just not yours. Product photography has to preserve the actual object: its shape, its logo, its colourway.

Instruction-based editing solves that. You pass the real photo and describe the change, and the model keeps the subject while replacing what you asked it to replace. No mask, no segmentation step in your pipeline.

Background replacement

The workhorse call. One photo in, one restaged photo out.

curl https://api.fastgencloud.com/v1/images/edits \
  -H "Authorization: Bearer $FASTGEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
      "model": "image-edit",
      "image": "https://cdn.example.com/products/mug-white-bg.jpg",
      "prompt": "place the mug on a polished walnut desk beside an open notebook, warm morning window light, shallow depth of field"
    }'

Fanning out a catalogue

Each background is its own job, so run them concurrently. Give every request a stable Idempotency-Key derived from the product and the scene — a retry then replays the original job instead of paying twice.

const SCENES = [
  "on a marble kitchen counter, soft daylight from the left",
  "on a linen tablecloth outdoors, dappled sunlight, summer",
  "on a dark slate surface, dramatic side lighting, product-hero style",
  "held in a person's hands, cosy interior, bokeh background",
];

const jobs = await Promise.all(
  SCENES.map((scene, i) =>
    fetch("https://api.fastgencloud.com/v1/images/edits", {
      method: "POST",
      headers: {
        Authorization: `Bearer ${process.env.FASTGEN_API_KEY}`,
        "Content-Type": "application/json",
        "Idempotency-Key": `mug-white-${i}`,
      },
      body: JSON.stringify({
        model: "image-edit",
        image: productUrl,
        prompt: `place the product ${scene}`,
      }),
    }).then((r) => r.json()),
  ),
);
// → four job ids; poll each, or pass webhook_url and collect callbacks

Prompt patterns that work

Edit models respond to concrete, imperative instructions. Name the surface, the light and the framing.

• Surface first: "on a polished walnut desk", "on brushed concrete", "on a linen tablecloth"
• Then light: "soft morning window light", "dramatic side lighting", "even studio softbox"
• Then optics if you care: "shallow depth of field", "product-hero framing", "45-degree angle"
• One change per request. Background plus colour plus angle in a single instruction usually gets you one of the three
• If the edit is too timid, raise steps to 20–30 and guidance to about 4 — the default is the fast distilled setting

Keeping the product honest

Generated imagery that misrepresents a product is a real commercial and legal risk, and it is also against the Acceptable Use Policy to present generated imagery as authentic where that misleads.

Practical rules: never let the model change the product itself, only its surroundings. Review before publishing. Disclose generated imagery where your market expects it — several jurisdictions now require it for advertising.

What it costs

Edits are billed per output image from a prepaid balance, so a four-scene fan-out costs four edits. A failed job refunds automatically. Current rates are on the pricing page — and because billing is prepaid, a bug in a batch loop stops at your balance rather than at your credit limit.