Business Baron needs a picture of a haulage yard in 1978, a picture of the same yard in 2012, a picture of a card shop, a picture of a card shop that is failing, and a picture of a market town high street that reads as Cumbria rather than Kent. Multiply that by more than two hundred business types, 91 towns and cities, and ten decades.
There are two of us. Commissioning that as illustration would cost more than the game will ever make, and drawing it ourselves would take longer than the rest of the game took to build. So we generated most of it, and we would rather say so plainly than let people assume otherwise.
There are about 1,100 illustrations in the shipped game.
One style definition, applied without exception
The thing that makes a set of images look like a game rather than a folder is consistency, and consistency at this volume cannot come from taste applied 1,100 times. It has to come from a definition.
So there is one style string. It specifies the medium, the palette, the lighting, the level of detail, the treatment of edges and the point of view, and it is prepended to every prompt without variation. The per-image part of the prompt describes only the subject: what the building is, what trade it houses, what decade it is, what the weather is doing, whether it is prospering.
Changing that string is a decision we take rarely, because changing it means regenerating everything. Twice we changed it anyway: once early, when the first attempt looked like stock photography, and once when we realised that the images with people in them dated badly and the images without people did not. The whole set was regenerated both times. That is only affordable because regeneration is a batch job rather than a commission.
The decade is the hardest attribute to hold. Signage, vehicles, shopfronts, street furniture and road markings are what tell you it is 1983 rather than 2003, and a generator that has learned a general idea of "old British high street" will happily mix them. The answer was to be specific in the prompt about the things that date a scene, and then to reject, rather than to fix, anything that came back wrong.
The budget is in dollars, and that is the point
Illustration is usually budgeted in weeks. Generated imagery is budgeted in dollars per image, and the number is small enough that the economics of the whole game change.
That does two things.
It makes rejection cheap. We generate several candidates per slot and keep one. The keep rate is well under half, and on the harder subjects it is much worse. A pipeline where throwing away nine images out of ten costs a few pounds is a pipeline where standards can be high. A pipeline where each image is a commission is one where you talk yourself into accepting the third attempt.
And it makes the set revisable. The images are not precious. If we decide in a year that the 1970s interiors are too clean, we regenerate the 1970s interiors.
The cost we did not expect is the reviewing. Someone still has to look at every candidate and decide. At several thousand candidates that is a real number of hours, and it is the actual bottleneck. We got it down by batching by subject, so that the eye is comparing twelve card shops against each other rather than a card shop against a haulage yard, and by writing down the rejection reasons so that the prompt could be fixed rather than the image replaced.
Shipping them is a separate engineering problem
A generated image arrives as a large PNG in a square aspect at whatever resolution the generator produces. None of that is what the game wants.
Every image is then cropped to the aspect the interface actually uses, resized to the sizes the interface actually displays rather than to one size that gets scaled at runtime, encoded to a modern format at a quality chosen per category, and checked.
The quality setting is not uniform. A full-width scene of a town can take more compression than a small portrait of a shopfront, because the artefacts hide in the detail. We settled the per-category settings by looking at them side by side at the size they are shown at, which is the only test that means anything.
The constraint that shaped the whole pipeline is download size. A game with 1,100 illustrations at careless settings is an enormous download, and a large download is a real cost to a player on a slow connection and a real cost to us in bandwidth. Getting the set to a sensible total was a week of work and worth more than any individual image.
Everything is reproducible. The prompt, the seed, the model, the crop and the encode settings for every shipped image are stored, so any image can be regenerated exactly, and the whole set can be rebuilt from the definitions. That is the same instinct that runs through the engine: if you cannot reproduce it from inputs you have kept, you do not really have it.
Some things still have to be drawn
Generated imagery is good at scenes and poor at anything that has to be exact.
The interface icons are drawn. They have to be legible at small sizes, consistent in weight, and unambiguous, and a generator produces something that looks like an icon rather than something that is one.
The maps are drawn, because the 91 towns and cities are real places and their relative positions and characters have to be right.
Anything containing text is drawn or set in type. Generated text in images is unreliable at best, and a shop sign with a nearly-word on it is worse than no sign.
The charts, the financial statements and everything else that displays a number are rendered by the interface from the engine's data, which is not artwork at all, and which is the part of the screen a player spends most of their time reading.
And the logo, the icon and the store artwork were made deliberately, because those are the things that represent the studio rather than the world inside the game.
Why we care about this
We would rather be straightforward about how the artwork was made than allow an impression we cannot support. Business Baron is a small studio's game, finished though not yet released, and the illustration budget is what let us spend the time on the ledger, the tax tables and the simulation instead. The art exists to make two hundred businesses in 91 places across nearly a century feel like specific businesses in specific places, and it is held to that test rather than to a portfolio's.