Generic Is the Default. Precision Takes Training

Ask a public AI model to generate an image of your product and it will hand you a product that isn’t yours. The proportions are off, the logo sits in the wrong spot, there’s a finish that doesn’t exist on anything you actually make. Ask it a second time and you get a different wrong product. Nothing here is broken. The model has seen millions of products in the general sense and it’s guessing at yours from an average that includes everyone else’s.

Foundation models are generalists, and generalists get vague the moment you ask for anything specific: a face, a product, the particular way a brand’s visual language actually looks. Those aren’t things the model knows. They’re things it approximates, and the approximation shows.

There’s a second question that gets less attention: what happens to the images you upload while you’re chasing that approximation. Midjourney’s terms give it the right to use your prompts and generated images to improve its models, and that applies even with stealth mode turned on to keep images out of the public gallery. OpenAI splits the difference by account tier. Free and Plus users’ content can end up training future models. Business, Team, and API customers get a contractual promise that it won’t. Upload an unreleased product shot for a quick mockup on the wrong tier, and it may not disappear after the session. It may become a data point in someone else’s model.

The fix for both problems is the same move: stop trying to get a generalist to know something specific, and train something narrow instead.

That’s a LoRA, in practical terms. A small trainable layer sits on top of a frozen foundation model and learns exactly the thing you need it to know, without retraining the whole model underneath it. Train one on a product and it stops guessing at proportions and finish, because it actually knows them. Train one on a face and it holds that face steady across angles and lighting instead of drifting toward a face like that one. Train one on a brand’s style guide and the output speaks that visual language instead of a decent impression of it.

The move that gets skipped is training several narrow ones instead of one broad one. A product LoRA, a face LoRA, a brand style LoRA, each shallow and specific on its own, combined only when a particular shot needs more than one. Same logic as anywhere else in production. A specialist who does one job exactly right outperforms a generalist doing five jobs approximately.

There’s a version of this most people haven’t thought about yet: photographer style. Working directly with a photographer, on their own body of work, with their own involvement in the process, a LoRA can be trained on the way they actually shoot: the lighting choices, the composition habits, the color decisions that make their work recognizable. It’s the photographer’s model, built from their own images, for them to use and offer on their own terms. This isn’t someone reverse-engineering a look from images scraped off the internet and slapping the photographer’s name on the result. It’s the photographer putting their own style into a tool they control.

That matters because AI-only content isn’t something most brands should be doing yet, for a long list of reasons that don’t need repeating here. But if a brand is going to do it anyway, the real choice isn’t between AI and a photographer. It’s between generic AI and a photographer who has trained their own style into the system and is offering it as part of their own work. One produces imagery that could belong to anyone. The other still carries a point of view, because the person who developed that point of view is the one who built the tool and is still the one behind it. If AI-only is happening regardless, that’s the version worth paying for.

None of this makes the training data question disappear on its own. It depends entirely on where the training happens. A LoRA trained on dedicated infrastructure, behind a private interface built around it so the brand or photographer runs their own model without ever routing it through someone else’s platform, keeps the data exactly where it started. A LoRA trained through a third party’s hosted tool just moves the same question one layer down, because their terms are still the ones that matter, however private the interface looks on the surface. The version worth having is the one where nothing about the process ever leaves a system built specifically for it.

Public models will keep getting better at the average case. Brands were never trying to make the average anything. The more interesting shift is smaller, more specific models, each built for one job and run somewhere private enough that the brand or photographer it was built for is the only one who ever touches it. Fewer generalists asked to guess at things they were never shown, and fewer shortcuts that quietly hand your material to a system that isn’t yours.