Writing a character's appearance down so it stays fixed
Where you can attach a reference, attach one — it is more reliable. This article is about the cases where you cannot: a character you are still designing, a look several people must share, batch work with no slot for an attachment.
Break the look into attributes
"A cute girl" returns something different every time. To reproduce a look, pin every attribute that would otherwise be left to interpretation.
- Hair: colour, length, parting, how the ends are treated
- Eyes: colour, shape, eyelid characteristics
- Face: outline, distinguishing marks such as moles or freckles
- Build: apparent height, shoulder width, apparent age
- Clothing: settle one default outfit and keep it
- Style: line weight, shading, texture
The key is passing the whole set every time. Anything you omit is handed back to the model's interpretation, and that is where drift starts.
Do I really repeat the whole description every time?
Yes. Anything you leave out goes back to the model, and that is where drift starts.
Separate what is fixed from what varies
Add the per-run instruction on top of the fixed block. Merging them into one paragraph loses track of what changed, and reproducibility goes with it.
| Class | Contents | Handling |
|---|---|---|
| Fixed | Hair, eyes, face, build, style | Passed verbatim every run; never edited |
| Semi-fixed | Outfit | Swapped per scene, from a settled list |
| Variable | Place, time of day, action, framing, light | Written for that run only |
Stop maintaining prose. Settle one image and attach it from then on.
Settle one image, then reference it
There is a more reliable endpoint than maintaining the text forever: generate candidates from the text, settle on one, and attach that image as the reference from then on.
Measured, a single reference held a subject's features across changes of angle, expression, and outfit. Treating the written description as a means of producing the first image — rather than a permanent system — makes the work stable.
Different models honour different instructions
The same sentence lands differently per model. These are the differences we measured.
| Instruction | GPT Image 2 | Nano Banana 2 |
|---|---|---|
| Camera angle (30°, 60°) | Weak — drifts back to frontal | Honoured |
| Multi-panel layout | Followed | Broke down |
| Text inside the image | Followed | Broke down |
For a character sheet with several angles side by side, this matters directly. Use Nano Banana 2 where the angle must land, GPT Image 2 where the arrangement must hold, and split the work across both.
Frequently asked
- Text or reference image — which is more reliable?
- The reference. Measured, one reference was more stable than describing the face. Treat written descriptions as the fallback when no reference exists.
- Do I have to repeat the whole description every run?
- Yes. Omitted attributes are left to the model, and drift starts there. Pass the fixed block in full each time.
- Should I describe the look when attaching a reference?
- Not the face. With a reference attached, describing the face lets the text take precedence and pulls the result away from it.
