Why she looks like a different woman in every photo
Character consistency is the hardest part of these products. Here is what is going wrong and what to look for.
You pick a character because of her face. Then you ask for a photo and get somebody else — same hair, same age, roughly the right vibe, different woman. It is the most common disappointment in this category and it has a specific cause.
Image models do not know who she is
A model generating a picture from your words has no memory of the character. Left alone it will draw a plausible person matching the description, and a different plausible person next time. Everything an app does to prevent that is a workaround.
The two common workarounds
The first is training a small model on many pictures of one face. Consistent, expensive, and slow to make, and it locks the character to whoever those training pictures were of.
The second is an identity adapter, which takes one reference photo and steers each new image toward that face. Cheap, instant, and it carries more from that reference than people expect — including, we learned the hard way, the expression. Give it one neutral portrait and you get a set of photographs where she is in exactly the same mood in all of them.
What to look for
- Ask for the same character in two different settings and compare the faces at full size, not as thumbnails. Thumbnails hide everything.
- Check the hands. Fingers are still the most reliable tell.
- Ask for something with a specific action and see whether the picture contains it, or only its setting.
- If she smiles the same way in every image, that is the adapter, not the model.
Why it is worth caring
Consistency is what turns a picture generator into a character. Without it you have a slot machine with a name attached, and after a dozen pulls you stop believing there is anyone there.