Bulk product image generation is not limited by how fast a model runs. It is limited by how fast a human can approve the output against the real product. A 200 SKU catalogue at four images each is 800 files, and every one of them needs an eye on it before it reaches a client. Generation takes minutes. Review takes days. Price the job on review time, not on credits.
Agencies get this backwards on the first big run. The pitch is that AI collapses a catalogue shoot into an afternoon, so the quote goes out at afternoon rates. Then someone opens folder three of eleven and finds the model has quietly changed a stitch colour on 40 files, and the margin is gone.
This guide covers what a bulk run actually costs, the throughput lever most teams never check, how to keep client work separate, and the three specific places batch generation falls over.
Key Takeaways
Review is the bottleneck, not generation. Plan a bulk run around approval throughput and the rest of the schedule follows. An image is 5 credits. 800 images is 4,000 credits, which is above Premium’s 2,500 monthly allocation and inside Ultra’s 8,000. Parallel generation runs 1, 1, 4, 8 and 16 by tier. Free and Basic run one job at a time. Ultra runs sixteen. On a catalogue that is the difference between a day and a week. Batch what is identical, not what is similar. A batch is only safe when every item in it takes the same treatment. Sample before you commit. Ten SKUs across the range tells you what 800 will look like. Separate clients at the brand profile level. One workspace with per-client brand profiles beats one workspace per client. Team seats start at Ultra. Five users included, extra seats $19 each. Below that every plan is single seat.
What bulk product image generation actually means
Three different jobs get called bulk work, and they have different failure modes.
| Job | What it is | Where it breaks |
|---|---|---|
| Volume | The same treatment across many SKUs | Products that need different treatment get it anyway |
| Variation | Many treatments of one SKU | Output drifts from the source product |
| Rescue | Fixing inconsistent inherited images | The source files are not good enough to fix |
Most agency catalogue work is volume plus rescue. A client hands over 200 supplier images shot by four different people, and the brief is to make them look like one catalogue. That is not a generation problem. It is a standardisation problem with a generation step in it.
Naming which job you are doing decides everything downstream, because the review criteria are different. Volume work is checked for consistency. Variation work is checked for fidelity to the product. Rescue work is checked for whether it improved anything at all.
The cost of a bulk run
DesignerBox charges 5 credits per image generate or edit. That makes the arithmetic simple and the plan choice obvious.
| Catalogue | Images per SKU | Total images | Credits | Smallest plan that covers it in one month |
|---|---|---|---|---|
| 20 SKUs | 4 | 80 | 400 | Basic, $15 |
| 50 SKUs | 4 | 200 | 1,000 | Pro, $35 |
| 100 SKUs | 4 | 400 | 2,000 | Premium, $75 |
| 200 SKUs | 4 | 800 | 4,000 | Ultra, $200 |
| 500 SKUs | 4 | 2,000 | 10,000 | Ultra plus credit packs |
Two things that table hides. Credits do not roll over into an unlimited pool, so a 500 SKU catalogue is a two month job on Ultra or a one month job with packs on top. And every failed generation costs the same as a successful one, so a 20% reject rate is a 20% cost increase, not a rounding error.
That reject rate is the number worth measuring on your first run. It is the difference between a quotable process and a hopeful one. Against a traditional shoot the comparison still favours this heavily, and what a product photoshoot actually costs has the day rates to compare against.
Commercial licensing starts at Pro, which is the floor for any client work regardless of volume.
Parallel generation is the throughput lever
This is the setting almost nobody checks before quoting, and on a catalogue it decides your calendar.
Parallel generations run 1, 1, 4, 8 and 16 across Free, Basic, Pro, Premium and Ultra. A Basic plan processes one image at a time. An Ultra plan processes sixteen.
Run the same 800 image job on both and the queue behaves completely differently. Sixteen concurrent jobs turn an overnight run into a lunch break. One concurrent job turns it into something you start on Monday and check on Thursday.
The practical rule for agencies: the tier decision on a catalogue job is a scheduling decision before it is a credit decision. Work out the deadline first, then pick the tier that clears it, then check whether the credits also cover it.
Batch what is identical, not what is similar
The most expensive mistake in bulk work is batching things that look like they belong together.
A batch is safe when every item in it takes exactly the same treatment. Same background, same light, same crop, same framing logic. The moment one item needs a judgement call, it is not part of that batch.
Sort a catalogue before you run it:
- Group by treatment, not by category. All items that need a white seamless packshot go together, whether they are mugs or shoes.
- Pull out the awkward shapes. Anything very tall, very wide, reflective, transparent or fine-detailed goes in a manual pile. Glass and jewellery are the usual offenders.
- Sample ten across the range. Two easy, two awkward, six typical. Run them, review them properly, and only then commit the rest.
- Fix the recipe, not the outputs. If eight of ten samples are wrong the same way, that is a brief problem. Adjust once and rerun rather than correcting eight files.
- Run the full batch in blocks. 50 to 100 at a time, reviewed before the next block starts, so a systematic error costs one block rather than the catalogue.
Step 3 is the one that pays for itself immediately. Ten samples is 50 credits. Discovering the same problem at 800 images is 4,000.
The heavy lifting on a real catalogue runs through the bulk catalog processor workflow, and for inherited supplier images specifically, supplier image rescue is the version built for mixed provenance. Teams driving this from Claude rather than the UI use the catalog batch processor skill.
Reviewing 800 images without reviewing 800 images
You cannot skip review, but you can make it cheap.
- Contact sheets, not files. Review 40 thumbnails on one screen. Consistency errors are visible at thumbnail scale, which is the whole point of the grid.
- Check the product, not the picture. Colour, texture, logo, proportion, material. Everything else is taste and the client will tell you.
- Spot check at full size. One in ten, plus every awkward item from your manual pile.
- Watch for systematic drift. Errors in bulk work are rarely random. If one is wrong, look for the other 39 that are wrong the same way.
- Keep the source photo beside the output. Fidelity is a comparison, not an impression.
That fourth point is what makes bulk review survivable. Individual errors are expensive to find. Systematic errors are cheap, because finding one finds all of them. Where AI product photo accuracy actually drifts covers what to look for first, and it is usually fabric and hardware rather than anything obvious.
Keeping clients separate
This is the agency objection, and it is usually phrased as “we already use several AI tools”. The honest answer is that the tools are not the problem. The organisation of the work is.
A stack of single-purpose tools produces client work that lives in whoever’s account generated it. Six months later a client asks for the same campaign on a new product, and the person who built it has left, the prompts are in a Slack thread, and the brand kit is a PDF someone emailed.
What changes that is not a better model. It is per-client brand profiles that hold the look, workflows that rerun the campaign for the next product, and one library everything lands in. Brand profiles is the mechanic, and running creative for multiple clients without drift covers the operational side.
Team seats are honest to state plainly: five users are included on Ultra at $200 a month, with extra seats at $19. Every plan below that is single seat. For a three person studio that is the real decision, and how agency creative pricing works out puts it against retainer maths.
The consolidation argument is worth checking rather than assuming, and the six tool stack does that arithmetic.
Where batch generation breaks
Three places, consistently.
The source images are not good enough. Rescue work has a floor. If the supplier photo is soft, small or shot under mixed light, a batch run produces 200 consistently mediocre images instead of 200 inconsistently mediocre ones. Check the worst 10% of the source set before quoting.
The product needs judgement. Reflective, transparent, very fine detail, or anything where the material is the product. These do not batch. Price them separately as manual work rather than discovering them mid-run.
The client has an unwritten rule. Every brand has one. The logo always faces left. The lid is always on. Shadows always fall right. It is never in the brief, it surfaces at review, and it invalidates the batch. Get it out of them during the sample stage, which is the real reason the ten sample step exists.
None of those are fixed by a better model, and it is worth being direct about that before a client expects otherwise. What they are fixed by is sequencing: sample, extract the rules, then run.
For the video equivalent of this problem, where file constraints add a second layer, TikTok Shop product videos and the 10 MB cap shows how a platform limit reshapes the whole production plan.
Getting started on a real catalogue
The shortest honest path:
- Pull 10 SKUs across the range and run them on the free plan, 112 credits, which covers 22 images.
- Review at full size, extract the client’s unwritten rules.
- Fix the recipe and save it as a workflow.
- Pick the tier by deadline, using the parallel generation numbers.
- Run in blocks of 50 to 100, reviewing between blocks.
What DesignerBox does for agencies covers the multi-client side, and Photo Studio is where the catalogue work itself happens. Scaling creative production without adding headcount is the wider version of this problem.
FAQ
How many product images can AI generate at once?
That depends on your plan’s parallel generation limit rather than on the model. DesignerBox runs 1, 1, 4, 8 and 16 concurrent generations across Free, Basic, Pro, Premium and Ultra. The total you can produce in a month is set by credits, at 5 credits per image.
What does it cost to generate 1,000 product images?
5,000 credits at 5 credits per image. That sits between Premium’s 2,500 monthly allocation and Ultra’s 8,000, so it is one month on Ultra or two on Premium. Budget for a reject rate on top, because failed generations cost the same as successful ones.
Is bulk AI product photography accurate enough for client work?
For volume and rescue work on straightforward products, yes, provided you start from the client’s own product photography and review the output. It is not reliable without review, and it is weak on reflective, transparent and fine-detailed products. Sample ten SKUs before committing to a catalogue.
How do agencies keep client brands separate?
With a brand profile per client rather than an account per client. The profile holds the look, and workflows saved against it rerun the same campaign for the next product. That is what stops the brand living in one person’s prompt history.
How long does a bulk product image run take?
Generation time is set by your parallel limit. Review time is set by your team, and it is the larger number. A useful planning figure is that reviewing at contact sheet scale with full-size spot checks runs several hundred images per person per day, so an 800 image catalogue is a multi-day approval job whatever the generation speed.
Do I need a commercial licence for client work?
Yes. Commercial licensing on DesignerBox starts at the Pro tier. Any images delivered to a client should be produced on Pro or above regardless of how small the run is.
Can I automate bulk product images without using the interface?
Yes. The MCP server exposes 43 tools across 8 groups, so Claude, ChatGPT or Cursor can drive the same models and apps directly. That is the route most teams take once a catalogue process is stable enough to script.
DesignerBox credit costs, plan allocations, parallel generation limits and seat counts taken from the product brief as of August 2026. Plan details change, so confirm on the pricing page before quoting a client. Individual results vary.
