Generating batch product images with AI and keeping them consistent

The core idea in one paragraph

A product catalog is not a collection of individual images. It is a set, and a set only works when every image looks like it came from the same shoot, the same studio, the same hand. Generating dozens of product images one at a time, with a fresh prompt each time, produces a catalog that looks scattered and unprofessional, because each image drifts to a different point in the style's range. The fix is batch thinking: one locked template that encodes the studio decisions, applied to every product with only the product description changing, followed by a consistency pass that normalizes the set. Get this right and you can produce a coherent catalog of dozens or hundreds of images at a fraction of the cost of a real shoot, with consistency that holds up against photography.

Why catalogs break when generated one at a time

Before the system, understand the failure, because the failure is subtle and expensive. When you generate each product image separately, even with similar prompts, the model lands each image at a slightly different point in the lighting, color, and finish range. Individually, each image looks fine. Side by side, they look like they came from different studios on different days, which destroys the professional read that a catalog depends on.

The drift comes from two sources. First, paraphrased prompts, where each image's prompt is similar but not identical, and the model treats every word difference as permission to vary. Second, the model's inherent output range, where even an identical prompt produces different images across seeds, because the model interprets within a range rather than at a point. A catalog generated without addressing both sources will always look inconsistent, no matter how carefully each individual prompt is written.

The implication is that catalog consistency is a systems problem, not a prompting problem. You cannot prompt your way to a consistent catalog by writing better individual prompts. You have to lock the shared decisions and generate the whole set within that locked system, which is a different way of working entirely.

The batch template

The foundation of catalog consistency is a single batch template that encodes every shared decision and leaves exactly one variable slot for the product itself. Everything in the template is locked: the background, the surface, the lighting, the lens, the depth of field, the style, the finish. Only the bracketed product slot changes between images. This is the catalog equivalent of the brand template, focused specifically on the photographic setup.

Here is a batch template structure, with the variable slot clearly marked:

[PRODUCT AND MATERIAL], centered on a seamless pure white background, large softbox light from above left with gentle fill from the right, soft contact shadow directly beneath, shot on a 100mm macro lens, deep focus with the entire product sharp, color accurate, studio product photography

Drop in a frosted glass perfume bottle with a brushed brass cap, then a matte ceramic mug in slate grey, then a pair of brushed aluminum headphones, and you get three products that look like they were shot in the same studio on the same day, because every decision except the product is identical. That visual unity is what makes a catalog feel professional, and it comes from the template, not from individual effort per image.

The seed strategy for batches

Seeds play a special role in batch generation, and using them well tightens consistency further. Recall from the consistency articles that a seed fixes the starting point in the model's latent space, and holding it constant keeps outputs in the same neighborhood. For a catalog, you have two viable seed strategies, each with a tradeoff.

The first is a single fixed seed across the whole catalog, which maximizes treatment consistency but risks making the set feel too uniform, more like clones than a family. The second is a small family of seeds, say three to five, rotated across the catalog, which preserves consistency while introducing enough variation that the set feels like a real shoot with natural minor differences. The second approach usually reads as more professional, because real catalogs have small natural variations, and a set that is too perfectly uniform can look artificial in its own way.

Whichever you choose, document the seed or seed family in the catalog spec, so the set is reproducible and so future additions to the catalog match the original batch. Undocumented seeds are a common reason later additions to a catalog look slightly off, because the new images used different seeds and drifted from the original treatment.

Handling product variants within the system

Real catalogs rarely have one image per product. A product comes in multiple colors, angles, or configurations, and each variant needs its own image that still matches the set. The batch system handles variants cleanly, as long as you treat them as part of the same locked treatment rather than as new images.

For color variants, keep the template identical and change only the color descriptor in the product slot. The same lighting, background, and lens apply, because the variant is the same product in a different color, not a different product. Generate each color variant with the same seed family as the parent, so the only thing that moves is the color, which is exactly what you want the viewer to compare across the variants.

For angle variants, the template stays the same but the camera position changes, which means adding or swapping the shot type and angle words while keeping the lighting and background constant. The key is that the lighting and background must not change with the angle, because a viewer comparing a front view and a side view expects the same studio setup, just from a different position. Lock the environment, vary only the camera, and the variants read as views of the same product rather than different images.

Document the variant rules in the catalog spec alongside the template and seeds, so variant generation is as repeatable as the primary generation. A mature catalog spec specifies not just how to shoot the hero image of each product, but how to shoot its variants, so the whole library stays coherent as it grows in both breadth and depth.

The generation checklist

Before generating a batch of any size, run through a short checklist that prevents the most common consistency failures. The checklist takes a minute and saves hours of rework.

1. Is the template finalized and frozen? Changing the template middle of the batch ruins the set. Lock it before the first image and do not change it.
2. Is the product list clean and consistent in format? Each product description should follow the same structure, so the variable slot reads predictably.
3. Are the seeds chosen and documented? Pick the seed strategy and record it, for reproducibility and for matching future additions.
4. Is there a reference image for the treatment? If the model supports style references, attach one canonical reference to every generation to anchor the treatment.
5. Is the output destination ratio decided? Generate in the final ratio to avoid cropping surprises.

Generating the batch

With the template locked and the checklist complete, generation becomes mechanical. For each product in the list, paste the template, swap in the product description, attach the reference if used, set the seed from your chosen strategy, and generate. Generate two or three candidates per product rather than one, because consistency is a numbers game and you want to select the image that best matches the set, not accept the first output.

As you generate, watch for outliers. Most images will land within the treatment range, but some will drift, especially for products with unusual shapes or materials that interact with the lighting differently. Flag the drifters for regeneration or for manual adjustment, rather than letting them into the set and breaking the consistency. The goal is a set where every image sits within the same narrow band, and culling the outliers is how you achieve it.

The consistency pass

Even with a locked template and careful generation, a batch will have minor inconsistencies that need a final processing step before it ships. Backgrounds may vary slightly in tone. Shadows may fall at slightly different angles. Color may drift warm or cool across the set. These small variations are invisible in isolation but visible when the catalog is viewed together, and they are what separate a good batch from a shippable catalog.

The consistency pass normalizes the set. Run the whole batch through a single processing step that unifies the background tone, aligns the shadow direction and intensity, and corrects color to a shared target. This is the last mile of catalog work, and it is where the set goes from a collection of good generations to a professional catalog. Tools that process a batch uniformly, applying the same normalization to every image, are ideal for this, because editing each image individually reintroduces the inconsistency you are trying to remove.

This is also the stage where any stray artifacts from generation get cleaned up. A watermark style reflection that slipped through, a stray highlight, a slightly misaligned crop, all of these get fixed in the consistency pass, so the final set is clean as well as coherent. Treat generation as producing the raw material and the consistency pass as producing the finished catalog, and your output will hold up against photography that cost ten times as much.

Scaling beyond dozens

The same system scales from a dozen products to hundreds or thousands, because the work per product does not grow. The template is written once. The seed strategy is decided once. The consistency pass is one operation applied to the whole set. Adding the thousandth product costs the same as adding the tenth, which is the entire economic advantage of AI generation over photography for catalogs.

At scale, the governance matters more, not less. A catalog of thousands generated without a locked system is a disaster, because the inconsistencies multiply across the set and become impossible to fix after the fact. A catalog of thousands generated within a locked system, with documented templates and seeds and a consistency pass, is a genuine competitive asset, because it gives the brand a complete, coherent, professional image library at a cost that photography could never match. The system is what makes scale possible, and scale is where the return on building the system becomes overwhelming.

The catalog as a living asset

A catalog is rarely finished. Products are added, updated, and retired over time, and each addition must match the existing set. This is why documentation matters so much. The template, the seed family, the reference image, and the consistency pass settings must all be recorded as part of the catalog spec, so that an image added six months later matches the original batch.

Treat the catalog as a living asset with a maintained spec, not as a single project. When a new product arrives, open the spec, apply the template, use the documented seed, run the consistency pass, and the new image drops into the set seamlessly. This is how professional catalogs stay coherent over years of growth, and it is entirely a function of the system you build once and maintain with discipline. Build the system, document it, and your catalog will scale and endure in a way that ad hoc generation never could.