Generated images have improved to the point where most subjects are indistinguishable from photographs, and hands and written text remain reliably wrong. Both failures come from the same property of how these models are trained.
The objective rewards plausibility, not correctness
Image models are trained to produce output that matches the statistical character of real images.
For most content that is sufficient, because a plausible cloud, fabric fold or tree is a correct one — there is no exact answer being missed.
Hands and text have exact answers. A hand has a specific number of fingers, and a word has a specific sequence of letters, and approximately right is simply wrong.
Counting is not something the architecture does
Producing five fingers requires maintaining a count across a region of the image, and these models have no counter.
They generate locally consistent structure, so each finger is well formed and the set does not resolve to the right number.
The same mechanism explains extra teeth, inconsistent numbers of buttons and railings with an implausible arrangement of posts.
Hands appear in training data in unhelpful form
Hands are small in most photographs, frequently blurred by motion, partly occluded, and shown at an enormous range of angles.
The model therefore sees a great many low-quality examples of an object with high geometric complexity, which is the worst combination for learning precise structure.
Faces, by contrast, are large, well lit, roughly frontal and abundant, which is why they were solved much earlier.
Text requires discrete symbols from a continuous process
Rendering a word means placing exact glyph shapes in an exact order, and the generation process works in continuous values.
What emerges is letter-like marks with the correct texture and typography, arranged into something that is not a word.
Models that handle text better usually do so because they were trained with a stronger text encoder and heavy emphasis on captioned signage, which improves short strings and still degrades on longer ones.
The remedies are external rather than architectural
Production workflows handle both problems by not asking the model to solve them: text is composited afterwards in a design tool, and hands are corrected by regenerating a masked region until one is right.
Pose conditioning, which supplies a skeleton as an additional input, improves hands substantially because it hands the model the structure it cannot count its way to.
Each of these treats the failure as a known limitation to route around, which is a more reliable position than waiting for the objective to start rewarding exactness it was never asked to produce.