This article answers a question nearly every seller asks: if I upload a photo of my own product, will the AI quietly "improve" it into a different one? Below is what the system does about that, when it can still drift, and what to do when it does.
Fidelity isn't a limitation, it's the floor for commerce assets
General-purpose AI image tools optimize for looking good. Commerce assets have to be the thing the buyer actually receives. A bottle whose curve changed, a label whose text got reflowed, a logo that migrated, a four-piece set that renders as three — however handsome the image, the goods don't match the listing, and marketplace review and buyer complaints are waiting at the other end.
So Xinsai writes product fidelity in as the platform's highest-priority global rule: the AI may change the background, the light, the scene and the model, and is required to leave your goods alone.
What the system does on every generation that includes your photo
Whenever a generation carries your own image as a reference, a fidelity constraint is force-injected at the front of the instruction sent to the model:
- product shape, proportions and packaging structure stay unchanged;
- the label and every word and graphic on it, the logo position and brand marks stay unchanged;
- color, material, distinctive details, accessories and item count stay unchanged;
- printed text on the packaging must not be removed, rewritten, translated or blurred — even when you ask for "no text in the image," that refers to marketing text and watermarks overlaid afterwards, not the product's own printed packaging;
- only background, lighting, camera angle, scene, motion and commercial presentation style may change, and only where you've explicitly asked for it.
None of that is optional and none of it is something you have to write into your prompt; it's added on every request. On top of that, fidelity-critical tools don't run on price-first model routing — the system pins a high-fidelity editing model, because cheap-but-loose models are exactly the ones most likely to repaint a real product into a lookalike.
Worth noting: pure text-to-image generation (no source photo at all) doesn't get this constraint, since there's no product of yours in the picture to preserve.
In rare cases it still drifts — here's the order to work through
Generative AI is inherently stochastic. A hard constraint pushes drift very low, but it can't make every single image perfect, and the more complex the product (transparent materials, dense fine print, multi-piece sets) the more room there is for error. Work through it in this order:
- Re-run it. A different roll of the dice often lands correctly. A re-run is a new generation and is billed at the normal rate.
- Switch to a steadier tool card. White-Background Product Image and Product Scene Image have shape, color, material, logo and label preservation written into their instructions. If you want zero risk at all, Hero Video animates your real photo compositionally — there are no AI-repainted pixels in the frame.
- Patch it with Fine-tune. The quick actions include *Fix Distortion* and *Remove Watermark & Text*, which edit a region of an existing result — faster than redoing the whole image.
- Get the product details right. Material and specs that a photo can't show become hard constraints the moment you type them in; see Why typing product details in beats letting the AI read the photo.
- Swap in a clearer source photo. With a blurry, dim or heavily occluded input, the model is guessing — and guessing is where drift comes from.
Failures refund automatically
When a job fails, the pre-authorized credits go back the way they came; and if a cheaper model ends up completing it, the difference is refunded too. Billing is CN¥1 = 10 credits, sign-up gives you 30 credits, and image generation starts at 2 credits — enough to put fidelity through a round of testing before you decide whether to pay anything. Details in Credits and pricing.
One distinction worth holding onto: "I re-ran it and still don't love it" isn't a failure and isn't refunded. Automatic refunds cover jobs that actually broke.