Can AI produce a character turnaround? Choosing models for a model sheet
A turnaround — front, side, and back arranged on one sheet — asks for two things at once: each view at the angle you specified, and all three staying aligned in one frame. Measured, those are the strengths of two different models.
The two requirements belong to different models
| Requirement | GPT Image 2 | Nano Banana 2 |
|---|---|---|
| Camera angle (30°, 60°) | Weak — drifts back to frontal | Honoured |
| Multi-panel alignment | Followed | Broke down |
| Labels and captions in-image | Followed | Broke down |
Is one run really not enough for a turnaround?
Get the angles right and the layout breaks; get the layout right and the angles soften.
Split the work
Rather than finishing in one image, generate each angle separately and assemble outside the model. That was the reliable route.
- Settle the character's appearance, in text or from a reference
- Generate front, side, and back separately on a model that honours angles
- Attach the settled image as a reference on every run so the subject stays the same
- Arrange and label in a layout tool, not in the generation
The third step is what makes it work. Generated independently, each angle comes back as a different person. Passing one settled image as the reference on every run changes the angle while holding the subject.
Use a model that holds identity
This route assumes a model that preserves features from a reference. One of the models we measured did not.
| Model | Identity retention | Angle instructions |
|---|---|---|
| GPT Image 2 | Held | Weak |
| Nano Banana 2 | Held conditionally | Honoured |
| Seedream 5.0 Lite | Did not hold | — |
Because the step needs both angle control and identity, expect to move between models — angles on Nano Banana 2, repairs on GPT Image 2. Not forcing one model to do everything finished faster.
Generate per angle, then attach the one settled image to every run. That holds the subject.
Keep labels out of the generation
Do not ask for "front", "side", or dimension figures inside the image. Even models that follow text instructions break down on small strings. Type is more reliable added afterwards.
Frequently asked
- Can I get a turnaround in one generation?
- Angle accuracy and layout accuracy belong to different models, so one run rarely satisfies both. Generate per angle and assemble outside.
- Won't separate angles produce different people?
- They will. Attach one settled image as the reference on every run, on a model that preserves features from a reference.
- Can labels go inside the image?
- Some models follow text instructions, but small strings break down. Add labels and dimensions after generation.
