How to make an AI product photo that matches your listing

The finished image should look like it was shot on a table with a soft light, not like it was generated. That is the whole test: if a buyer can tell, the image is not doing its job.

This tutorial uses one real photo of the product and changes only the scene around it. The method works with any tool that accepts a reference image; the screenshots here are from Aurora Canvas, which has a free tier large enough to finish this without paying.

What you need before starting

A phone photo is enough, but it has to be honest: daylight, no flash, product in focus, plain background. Retouched reference photos produce retouched-looking output, and the model has no way to know which flaws you intended to keep.

If the product has fine text on it, shoot closer than you think you need to. Label legibility in the reference is the single biggest factor in whether the label survives generation.

Why the prompt describes the scene only

Every failure in product generation traces back to the same mistake: describing the product. The moment you write “a white ceramic candle jar”, the model starts drawing its own jar and your reference becomes a suggestion.

Describe light, surface, shadow and framing. Those four variables are enough to move a product from a kitchen table to a studio, and they leave the object alone.

Step by step

1. Upload the reference and crop tightly, so the product edge touches the frame on all sides.
2. Paste the scene template, replacing only the bracketed parts.
3. Generate four variants rather than one.
4. Pick the variant with the most believable shadow, then fix the shadow: 3-5% grain, slight edge blur.
5. Export layers at 2x, then crop to 1:1 and 4:5.

The three failures worth knowing

The label warps. Almost always a loose crop. Re-crop and regenerate; prompt changes will not fix it.

The product floats. The prompt asked for a long shadow but the reference has no ground contact. Add “contact shadow directly beneath the product” and regenerate.

The colour drifts. You described the product colour somewhere in the prompt. Remove every colour word that refers to the product itself and keep only light temperature.

When not to use this

If the product is transparent, reflective or has moving parts, a generated scene will fight you and a real lightbox is faster. Glass needs caustics, and most models render it as plastic on the first try.

For the workflow that pairs this with a local text model — writing listing copy from the same product spec — see the Nimbus 7B local setup.

Keep the product exactly as in the reference: same shape, same label, same material.

Scene: [surface — e.g. warm oak table, matte concrete, linen cloth]
Light: [lighting — e.g. soft window light from the left, single studio softbox, late afternoon sun]
Shadow: [shadow — e.g. soft contact shadow to the right, long low shadow]
Framing: [framing — e.g. three-quarter view, eye level, product centred with space above]
Style: clean commercial photography, natural colour, no text overlays, no props touching the product

extra text, watermark, logo changes, altered label, floating product, hard black shadow, oversaturated colour, plastic skin, multiple products

Same scene, change only the light direction. Four variants: left, right, top-left, backlit rim.