# ANOMALY Knowledge · GPT Image 2 reference

**One job: make the master character image.** GPT Image 2 is the best model here at a believable single human face and at clean shapes and text, so it creates the character once. Nano Banana Pro then reuses that character in every shot forever. Do not use GPT Image 2 for build clips or frame recreations.

Also good for: a clean logo, and tight product close-ups where no character appears.

---

## The master character prompt

This is the most important image you will generate, because every later clip inherits it. Spend the effort here.

```
Casual iPhone photo of [age range] [gender] [ethnicity], [hair: exact style and how it sits],
[skin tone] with visible pores and natural texture, [face shape], [one signature accessory].
Wearing [outfit]. Standing in [ordinary real setting], soft natural window light.
Neutral relaxed expression, looking slightly off camera.
Shot on iPhone, handheld, natural light, real person, organic content.
9:16 vertical, waist-up framing, plain uncluttered background.
No studio lighting, no perfect skin, no CGI, no ring light, no posed model,
no makeup gloss, no watermark, no text.
```

**Generate 4, pick 1.** Outputs vary a lot; one generation is a coin flip. Then never generate the character again — that single approved image becomes the reference you pass to Nano Banana Pro for the rest of the product's life.

---

## Why "not model-perfect" is a conversion rule, not taste

Never write *attractive*, *beautiful*, *gorgeous*, *flawless* or *model*. Write **real, slightly imperfect, authentic**. A model-perfect face reads as an advertisement, and the entire mechanic of AI-organic content is that the viewer believes they are watching a person, not an ad.

Concretely useful imperfection cues: visible pores, a slightly uneven smile, flyaway hair, a plain unstyled top, a lived-in room behind them.

---

## Character DNA — write it once, here

The moment you approve the character image, freeze the description as a fixed block of words:

```
[age range] [gender] [ethnicity], [hair, very specific], [skin tone + texture detail],
[face shape], [one signature accessory]
```

Paste that block, unchanged and unparaphrased, into every prompt you ever write for this character. Rewording it between clips is the single biggest cause of the face drifting.

---

## Settings

| Setting | Value |
|---|---|
| Aspect ratio | 9:16, stated in the prompt text *and* the parameter |
| Resolution | 2K for the master character. 1K only for throwaway drafts |
| Count | Generate 4, pick 1 |
| Reference image | None needed for the master character — you are creating it |

---

## Logos

GPT Image 2 handles clean shapes and short text better than the alternatives, so it is the right tool for a simple brand mark:

```
Simple flat vector logo for [brand name], [one visual motif].
Two colours maximum, thick even strokes, high contrast, centred, plenty of negative space.
Plain solid background. No gradients, no 3D, no photorealism, no drop shadow, no mockup.
```

Keep it simple. A logo that looks like a big corporate brand is off-key for an organic account.

---

## Failure modes

| Symptom | Fix |
|---|---|
| Face looks like an ad | Remove any beauty adjectives; add pores, texture, a neutral expression |
| Background too clean | Ask for an ordinary lived-in room with clutter |
| Skin plastic | Add "visible pores, minor blemishes, realistic skin texture" |
| Text in a logo is garbled | Shorten the text, or leave the text out and add it in the editor |

**When the same fault keeps coming back**, do not re-roll. Re-read your own prompt and find the word pulling the wrong output — that is nearly always where the fault is, not in the generation.