Editing models remove the background from your real photograph rather than generating a new picture that resembles it. That distinction matters most in a catalogue, where a product photo that is not the product becomes a returns problem. The run takes about two and a half hours and roughly twenty minutes of your attention.
An online seller. Three years of catalogue shot by three different photographers. Grey backgrounds, beige backgrounds, one set shot on a kitchen table. This is one of four jobs worked through in what AI can do that a chat window cannot, and it is the one that looks least like an AI problem until you try it.
The goal is boring and specific: every product on clean white, cropped the same way, sharp at 2,000 pixels, sized for the storefront. Nothing creative. Just consistent.
Why the chat window stalls
Two reasons, and the second one is the one that surprises people.
The arithmetic. At ninety seconds of your attention per photo — describe it, look at the result, decide, download — 2,400 photos is about sixty hours. That is a month and a half of evenings for something with no creative decisions in it.
The generation problem. Chat image tools make pictures. You upload a photograph, describe what you want, and get back an image that looks like your product. Close, but redrawn: the logo has drifted, the stitching is in a different place, the shade of blue is not quite the shade of blue you sell.
For a mood board, close enough is fine. For a catalogue, a product photo that is not the product is a returns problem, a review problem, and in some categories a compliance problem. The whole job depends on the pixels of your actual product surviving intact.
The models that do the work instead
Three models, each doing one thing, in order. None of them invents anything.
BiRefNet cuts the subject out of the background pixel by pixel. It is a segmentation model — its entire job is deciding which pixels are the product and which are the kitchen table. It does not draw a new product. It isolates the one already there, edge hair and all.
Real-ESRGAN sharpens and enlarges the older, lower-resolution shots so the whole catalogue holds up at the same size. The 2019 photographs were shot smaller than the 2024 ones, and a storefront grid makes that obvious.
Qwen-Image-Edit handles only the stubborn ones — evening out harsh lighting, or a shadow the cutout could not save. It edits the real photo rather than inventing a new one, which is the same principle applied to the hard cases.
The order matters. Cutting out first and sharpening second means the upscaler never wastes work on background it is about to throw away.
The run, step by step
- STEP 1Load the folder4 min
- STEP 2Cut out subjects1 hr 10 min
- STEP 3Crop and place on white12 min
- STEP 4Sharpen and resize55 min
- STEP 5Export the set8 min
Two of those five steps are almost all of the time. Cutting out subjects and sharpening the older shots are the work; loading, cropping and exporting are rounding. That shape is typical, and it is why "how long does it take" and "how long do I have to be there" have such different answers.
Your part of it is ninety seconds at the start: say what you want in plain words, point us at the photos, and confirm the cost before anything runs. The twenty minutes in the ledger is spot-checking a preview grid afterwards, not supervision.
What you will have when it is done
The full set, renamed and numbered, ready to upload. A preview grid so you can check the work before downloading anything. And the handful it could not save, flagged separately with a reason. A product shot against a background the same colour as itself will defeat any cutout, and it is better to be told which ones than to find out live on the storefront.
That last item is worth dwelling on. A batch job that quietly returns 2,400 files, some of them wrong, is worse than one that returns 2,380 and a list. Silent failure on a pile this size is undetectable until a customer finds it for you, so anything the models were not confident about comes back separated rather than mixed in.
Why the older photos are the expensive part
The 2019 shots are the reason this job is two and a half hours rather than one. Not because upscaling is slow in itself. Low-resolution source material makes every earlier step harder: edges are softer, so the cutout has less to work with, and compression artefacts around a product look enough like the product to confuse a segmentation model.
This is the general shape of batch work. The average item is quick and the worst tenth sets the schedule. It is also why the first step of the run sorts the files before any cutting starts. Sorting first means the difficult photos get the slower treatment and the easy ones do not pay for it.
If most of your catalogue was shot on the same camera in the same year, this job gets meaningfully faster and cheaper. Mixed-vintage catalogues are the common case, though, which is why the numbers here assume one.
The one decision worth understanding
This job is the clearest example of the first question in the model-selection framework: are you making something new, or changing something that exists?
It looks like an image job, so people reach for image generators. It is an editing job, and editing models are a different family entirely — trained to preserve rather than invent. That single distinction is the difference between a usable catalogue and 2,400 pictures of products you do not sell.
The same fork appears in the short-video job, where the answer is to start every clip from a real product photograph rather than from a paragraph describing one.
What it costs to be wrong
Worth naming, because the failure mode here is quiet. A generated catalogue does not look broken. It looks fine, ships, and then produces returns and a slow trickle of reviews saying the item did not match the picture. By the time that is legible in the numbers, six months of listings need redoing.
Editing models cost the same and carry none of that risk, which is why the decision is not really a trade-off. It is just a fork people take the wrong way because both sides are called AI.
Once you have seen what this looks like end to end, how it works shows the same job as a handover: your three steps against the five that run afterwards.