The tenth AI image looks great. The five-hundredth looks like a different company made it. Generating product imagery at scale is easy; keeping it recognisably yours — and keeping the product itself truthful — is the actual engineering problem. Here is the system we build for Vietnamese sellers who need hundreds of on-brand images a month without a studio.
The golden rule: never let AI redraw the product
Everything else in this article is negotiable. This is not:
In practice: shoot each product once, well, on a neutral background; cut a clean mask; and let the pipeline place that masked photo into as many generated scenes as marketing wants. The product pixels never change.
A style guide the pipeline can execute
A brand book PDF is for humans. For the pipeline, express the style guide as reusable prompt fragments — short, named blocks that are concatenated into every generation request:
- Scene fragment — the world your brand lives in: "minimal bathroom shelf, morning light, soft shadows" versus "vibrant Tết market stall, warm lanterns".
- Palette fragment — your brand colours named explicitly, plus what to avoid. Palette locking works far better as a constraint stated in every prompt than as a hope.
- Camera fragment — focal length feel, angle, depth of field, so a grid of thumbnails reads as one photographer's work.
- Negative fragment — the standing list of things that must never appear: competing logos, text artefacts, hands touching the product.
Version these fragments like code. When the brand refreshes, you edit four fragments — not five hundred prompts scattered across chat histories.
Consistency techniques that actually work
Seeds and reference images
Fixing the random seed while varying one prompt element gives you controlled variation — same scene, different angle — instead of a lottery. Reference-image conditioning (feeding an approved output back in as a style anchor) keeps a campaign's backgrounds in one visual family. Combine both: a small library of approved "anchor" images per campaign, each with its seed and fragment set recorded, so any image can be regenerated or extended months later.
The compositing and finishing chain
After generation, every image runs the same finishing pipeline: composite the masked product photo, match lighting and shadow direction between layers, apply the palette check, upscale to the largest size any channel needs, then export per-channel derivatives. Upscale once at the end — upscaling before compositing softens the product edges you worked to keep sharp.
Asset management: the unglamorous half
Marketing cannot use images it cannot find. Enforce a naming convention that encodes product, scene, campaign, version and ratio, and store fragments + seed alongside each asset as metadata. Every published image also gets descriptive alt text — written once by the pipeline from the product data — which is both an accessibility requirement and free SEO for your product pages. The same naming discipline feeds your video pipeline and your auto-posting calendar, so one asset ID traces from generation to post to ad metric.
The traps
| Trap | How it bites | Defence |
|---|---|---|
| Trademark elements sneaking in | Generated scenes include recognisable third-party logos, characters or trade dress | Negative fragments + an automated logo/brand-detection pass + human review of new templates |
| Inconsistent lighting between layers | Product lit from the left composited onto a background lit from the right — the eye reads it as fake instantly | Shoot products under neutral, diffuse light; state light direction in the scene fragment; relight or shadow-match in finishing |
| Image weight | 4 MB heroes tank page speed and conversion, especially on mobile data | Compress heroes to under 200 KB in the export step — modern formats make this routine |
| Untracked generations | Nobody can reproduce or extend last month's campaign look | Store seed, fragments and model version with every asset |
Where this sits in the bigger machine
Imagery is one station in the marketing factory of the end-to-end automation stack: product data in, on-brand assets out, feeding listings, ads and the posting calendar. Get the golden rule and the fragment system right first — scale without them just produces off-brand images faster.
A practical rollout order: photograph and mask your top twenty products, write the four fragments with whoever owns the brand, generate one campaign's worth of anchors and get them approved, then open the pipeline to the full catalogue. Two weeks of setup buys months of one-click campaign imagery — and every image is defensible, because the product in it is the product the customer will actually receive.