What AI images are good at, and where they fail

Generated images are very good at anything where nobody will measure the picture against a real object, such as worlds, mood, light, weather, places and moments that do not exist yet. They are weak wherever somebody will. That single test sorts almost every job you will be offered, and it is worth learning before any technique, because it decides whether a brief should be generated, photographed or built from both.

The first strength is worlds. The set, the scenery, the season or the interior around a product can be shown in minutes, well enough for a room of people to agree on it, long before anything is built or booked. These are concept frames, images made to settle an argument rather than to be published. A campaign that cannot decide between a pale oak Scandinavian kitchen, a 1970s kitchen in brown tile and a small Mumbai apartment kitchen can see all three by lunchtime, and choose one before anybody pays a location fee or a set builder. Nobody publishes those frames, but the decision they produced shapes everything that follows.

The second strength is previsualisation, which film crews and advertising agencies practised for decades with storyboards and rough animatics before any of this existed. A shot list written as six frames, with the light, the surface and the camera height decided, can be generated for well under a dollar and laid out on one page. Walking into a pitch with the campaign already visible removes the part of the process where a client approves words and then dislikes pictures. The shoot still happens. It happens against an agreed image, and the changes people would have made on the day, to the crop, the props or the time of day, have already been made for nothing.

The third is the unglamorous middle of a real production. The jobs are extending a background, filling a plate, building a surface and replacing the sky that was grey on the day, and none of them is the hero image, yet all of them used to take a retoucher hours. Film has done this since 1907. That year Norman Dawn painted crumbling California missions on glass for a film called Missions of California, the first known matte painting shot, and set extension has been part of the craft ever since. Growing a real photograph outwards is now one of the most reliable things a model does, because the real photograph has already fixed the light and the perspective before the model adds anything.

The fourth is volume, once an idea is agreed. Forty versions of an approved image, cut for a story, a banner, a poster and a product page, in different markets and seasons, used to be forty days of work and can now be an afternoon of generating and an evening of choosing. The order matters here. Volume before an agreed idea gives you forty versions of nothing in particular, and choosing among them becomes the whole job; volume after it gives you a family that holds together, because each image answers the same brief.

The weak list is just as stable. A later lesson takes it apart in detail, but you should know its shape from the start. Generated images fail at exact reproduction of a specific object, which in commercial work means your client's actual product, and they fail at small legible text, at hands where they touch something, at reflections that agree with the room, at counting and at holding the same face from one picture to the next. Every item on that list is something a viewer can check against reality. So the rule this course works to is that the thing being sold is photographed or modelled, then composited into a generated world, and only the world is invented.