You make a character you like on Monday. On Thursday you open the same tool, describe her again, and get her cousin.
The reflex is to blame the model. Most of the time the model isn’t the reason, because the tool was never built to remember anything.
Most tools have no concept of a character
Look at what a typical generator keeps between sessions: your prompts, your outputs, your billing. Not a character. Nothing in the system says “this is Maya, here’s her canonical face, use it every time”, because a character isn’t an object in that system. It’s a description you happened to type twice.
Which is why the good result is so hard to get back. You didn’t create a character on Monday, you got a lucky sample, and by Thursday you’re describing that sample from memory to a system that never kept it.
The tools that hold identity well are the ones where a character is something you make once, with its own record, that every generation is required to consult. The rest is a prompt box with history attached.
A generator optimises for the image, a diary needs the run
The unit of value differs, and the architecture follows the unit.
An image generator is judged on the picture in front of you. Every product decision points at making this one output impressive: better lighting, better composition, more detail, a stronger style. None of that has anything to say about whether next week’s output matches.
A diary comic is judged across the run. Strip 40 has to sit beside strip 1 and read as the same person. The individual images can be far simpler than a generator’s best work, and the job they’re doing is harder, because it’s a constraint over time rather than a quality inside a frame.
Very little in the tooling landscape was built for the second job. A tool can be excellent at the first and useless for this.
Consistency is invisible in a demo
The commercial reason is simpler and probably decisive.
A demo shows you three beautiful images. Drift doesn’t appear in three images. It appears at thirty, spread over weeks, by which point the user has either subscribed or gone. Nobody churns on day two over a changed face, because on day two there’s nothing to compare it against.
So the expensive feature is the one that can’t be shown where the buying decision happens. Add that a persistent character usually means storing references or training something per user, which costs money per user per month rather than per image, and the incentives line up against it twice over.
Photo pipelines forget by design
Then there’s the category that turns your selfie into a cartoon.
Those start from an image every time, which sounds like it ought to guarantee consistency and does the opposite. Different photo, different light, different angle, different expression, different haircut: the output tracks the input, and the input is a new photograph each time. Your cartoon self is being re-derived nightly from whatever you happened to look like in that picture.
For a profile picture, fine. For a Wednesday strip that has to match Tuesday’s, it’s a moving target wearing the costume of a stable one.
What a tool that remembers has to have
Four things, and you can check for all of them from the outside.
A character record that outlives the session. Made once, editable, referred to by name rather than re-described.
Canonical references the system keeps. Not your description of the character. Images the tool holds and conditions on.
Regeneration against those references, every time, including when the pose, scene and expression all change at once.
A way to correct drift. When a face comes out wrong you need to tell it so, and have that feed back into the record, rather than rerolling until something looks close.
A tool with none of those will forget you, and no amount of prompt discipline at your end fixes an architecture with nowhere to put the answer.
Before you put a year into one
Make the character. Wait a day. Open the tool cold, with no history in the window, and get them back.
That’s the whole test, it costs one day, and it maps onto how a daily habit will use the thing. A tool that passes has built something real underneath. A tool that fails will go on producing lovely pictures of strangers, and in a year you’ll have an archive of people you don’t know.