How to attach a reference image: count, quality, and what not to write
You attached a reference and got a stranger back. There are three causes: the model, the image you chose, and the extra sentences in your prompt. In that order.
1. Model selection comes first
Accepting a reference is not the same as holding a likeness. In our tests one model returned a different person every time regardless of the reference. Prompting does not fix this.
| Model | Identity retention | Reference cap |
|---|---|---|
| GPT Image 2 | Held | 16 |
| Nano Banana 2 | Held conditionally | 14 |
| Seedream 5.0 Lite | Did not hold | 14 |
Surely filling the cap makes it closer?
Filling it averaged the features out. If you attach several, vary the angle.
2. One good image beats several
Measured, a single clear image of the face was more stable than attaching more. Filling the cap tends to average the features away.
- The face is clearly visible
- Low contrast between light and shadow — no half of the face crushed
- Not an extreme angle
- Enough resolution that compression artefacts are not visible
If you do attach several, vary the angle. Near-identical shots from the same angle add no information.
The reference decides who. The text decides what is happening. Do both and they fight.
3. Do not describe the face
This mattered most. With a reference attached, adding "almond eyes" or "a straight nose" pulls the result away from it, because the text takes precedence.
- Write: place, time of day, clothing, action, framing, light
- Do not write: facial features, hairstyle (unless changing it), age, build
Let the reference specify who and the text specify what is happening. Specifying the face in both puts them in competition.
Respect per-model limits
The same reference behaves differently across models because the accepted settings differ. A shared UI listing one set of controls will produce combinations that fail somewhere.
| GPT Image 2 | Nano Banana 2 | Seedream 5.0 Lite | |
|---|---|---|---|
| Images per run | 1–10 | 1 only | 1–4 |
| Resolution | Not supported | 512 / 1K / 2K / 4K | 2K / 4K (1K returns 400) |
| Quality | auto / low / medium / high | Not supported | Not supported |
Frequently asked
- How many reference images should I attach?
- One clear image of the face was most stable. If you attach several, vary the angle rather than repeating similar shots.
- Why doesn't it look like my reference?
- Usually either the model cannot hold identity, or your prompt describes the face. The first can only be fixed by changing models.
- How high does the reference resolution need to be?
- Enough that the face is clearly visible, not crushed by shadow, and free of obvious compression artefacts. Avoid extreme angles.
