How to fix hands, text and small details in AI images
Why AI images still get hands, lettering and small objects wrong, what to change in the prompt, and when a re-roll or an edit model is the faster fix.
The picture is almost right. The light is good, the composition works, and then you notice the sixth finger, or the shop sign that reads like a spelling test gone wrong. Small details are where image models still slip, and they are also where a picture gets thrown away. Most of these slips have a known cause and a cheap fix, so here is a troubleshooting list rather than a theory lesson.
Why the small things go wrong
An image model does not count fingers or spell words. It paints texture that looks like hands and looks like letters, at a scale where a few pixels decide whether a thumb is a thumb. The newer models in the catalog are far better at both than they were a year ago, but the rule still holds: the smaller a detail is in the frame, the more likely it is to be wrong. Every fix below is some way of making the detail bigger, simpler or less important.
Hands: give them less to do
- Name the pose. "Hands resting on a table" or "holding a coffee cup with both hands" is a pose the model has seen thousands of times. "Gesturing expressively" invites fingers to go their own way.
- Keep hands large or keep them out. A close crop where the hands fill part of the frame renders better than tiny hands at the edge of a wide shot. If the hands do not matter, frame them out: "framed from the chest up".
- Avoid tangles. Interlaced fingers, two people holding hands, a hand gripping a thin object: these are the hardest cases. Simplify to one hand, one object.
- Look before you fix. On many renders the hands are fine. Check at full size before you spend anything on a correction.
Close-up portrait of a potter at a wheel, both hands shaping wet clay, hands large in the frame, soft window light from the left, shallow depth of fieldText: short, quoted, and big
Lettering is the detail models improved most, and still the one to treat with care. Three rules cover most of it:
- Put the exact words in quotes: a sign that reads "OPEN LATE". The quotes tell the model these are the letters, not a description.
- Keep it to one to four words. A title renders; a paragraph turns into letter-shaped noise.
- Give the text a surface and a size: "large white letters painted on a brick wall" beats "with text".
If the words matter to the finished piece - a poster headline, a price, a brand name - the honest answer is often to render the picture with a clean space for the text and set the lettering yourself in a design tool afterwards. It is sharper, it is spelled right, and you can change it without rendering again.
Small objects and counts
Models are unreliable with exact numbers and with many small things touching each other. "Exactly three candles" may come back as two or five; a table full of cutlery may melt into one silver shape. Ask for fewer, larger items, and describe their arrangement: "three candles in a row on the left side of the table" gives the count a place to live.
Re-roll or edit?
When a detail is wrong you have two options, and the cheaper one depends on how good the rest of the picture is.
- Re-roll when the whole picture is only fine. The same prompt gives a new picture, and the next hand may be perfect. Drafting on Still Lite, from €0.30, keeps re-rolls cheap.
- Edit when the picture is right and one detail is not. Several models in the catalog take your image as input and follow an instruction such as "fix the left hand so it has five fingers, change nothing else". Look for image-to-image models on the models page; the price shows in the composer before you run it.
What the studio does not have is a brush for masking one area by hand. An edit model decides what to change from your words, so keep the instruction short and say what must stay the same. If it keeps touching the rest of the picture, re-roll instead.
A quick checklist before you render
- Are the hands doing a common, simple thing, and are they big enough to render well?
- Is every piece of text in quotes and four words or fewer?
- Did I ask for an exact number of small things? Could it be fewer and larger?
- Would setting the text myself later be better than asking the model for it?
Prompt structure matters as much as these fixes; the guide to writing image prompts covers what goes where. For a character who has to look the same across several pictures, reference images do more than any wording can. When you are ready, open the studio and try the checklist on your next picture.
Questions
Why do AI images get hands wrong?
Hands are small, have many joints and take thousands of poses, and the model paints them as texture rather than counting fingers. Simple poses and hands that are large in the frame render much better.
Can AI image models write readable text?
Newer models handle short text well. Keep the words in quotes and to four or fewer, and for anything important, leave space and set the text yourself afterwards.
Should I re-roll or edit a picture with one bad detail?
Re-roll if the rest of the picture is only average. If it is right apart from one detail, send it to an image-to-image model with a short instruction that says what to fix and what to keep.
