Pillar guide

AI E-Commerce Product Photography: The Complete 2026 Workflow

2026年10月 · minimaxi.design 電商視覺實驗室

Most product photos that sell were not shot in a studio. They started as a phone snapshot or a supplier image, and were then cut out, relit, cleaned up and resized until they met the marketplace's rules. AI image editing has made each of those steps cheap enough to run on every SKU — as long as the product itself does not change. A model that "improves" your logo, smooths the fabric weave or re-spells the label has produced an image you cannot legally list.

This guide is the end-to-end workflow we recommend: which step uses which model, which prompts keep the SKU locked, and what each step costs per image on this site.

The five-step pipeline

1. Clean cutout. Separate the product from its background first, on background removal. Every later step is easier — and safer for the product's geometry — when the model only has to paint around a fixed object.
2. Scene or white. For the main image, place the cutout on pure white (RGB 255,255,255). For secondary images, generate a scene around it with product relighting, which matches light direction and adds a contact shadow.
3. Precision fixes. Dust, scratches, a wrong colour variant, an outdated label line: use inpaint with a mask so only those pixels are regenerated.
4. Upscale last. Run upscale on the final composite, not on the raw input, so the upscaler is not asked to invent detail that a later edit will overwrite.
5. Set and video. Assemble the listing set in Marketing Studio and, if the channel rewards motion, turn the hero still into a short clip with product video.

Which model does which job

Two model families do almost all of the work. Flux Kontext / Flux 2 (Black Forest Labs) is strongest at scene-level changes: new backgrounds, new light, new context, while keeping a referenced object consistent. Qwen Image Edit (Alibaba) is strongest at precise, local and text-aware edits: changing words on a box, removing a sticker, recolouring one part. Seedream is useful for generating consistent variation sets. The detailed comparison is in Flux vs Qwen Image Edit.

本站 65 個按張計費的修圖與圖片工具模型，定價從每張 $0.0060 到 $0.22。沒有月費訂閱，按產出張數付費。

Prompts that keep the product locked

Edit models obey constraints you state explicitly and drift on the ones you do not. A reliable product-edit prompt has four parts: the change, the scene, the light, and the lock list.

Place the bottle on a light oak table beside a window, soft morning light from the left, shallow depth of field. Keep the bottle, label text, logo, cap and colour exactly unchanged. Do not alter the product's shape or proportions.

Name the parts that must not change — label, logo, stitching, colourway — every time. "Keep the product the same" is weaker than listing the four things a shopper would notice.

What a listing actually costs

At the live list price of Qwen Image Edit Plus ($0.032 per image), a full eight-image listing set is about $0.25 if every image works first time. With Flux Kontext dev ($0.038 per image) the same set is about $0.30. Budget for retries: three attempts on a difficult hero shot is $0.095 on Qwen. For a catalogue of 100 SKUs × 8 images that is roughly $25.20 on Qwen or $30.00 on Flux Kontext — with no monthly seat.

Quality checklist before you upload

• Zoom to 100% on the label: every character matches the physical product.
• Logo edges are crisp, not re-drawn; colours match the real SKU under daylight.
• Shadows fall in the same direction as the scene's light.
• The main image meets the marketplace rule set (see Amazon white background).
• Nothing is shown that is not in the box — props in lifestyle shots are fine, but not on the main image.