How to Use GPT Image 2 for Anime Key Visuals

How to Use GPT Image 2 for Anime Key Visuals

This guide teaches the model through a world. Instead of a list of magic phrases, you get five lessons — a bridge, a classroom, a pair of hands, a light and an absence — each built from a character card, an event and a camera position. By the end you should be able to write an anime key visual prompt in about four minutes and repair it in one pass.

It assumes no prior experience with image models and no encyclopaedic knowledge of A Silent Voice. You need the world’s cards, GPT Image 2 and one decision per lesson: what is happening in the next ten seconds. Every example image below is a real published work from this world.

What you need before you start

Lesson 1 — Read the card before the prompt

A prompt is not a description of a picture. It is a set of instructions for staging a scene that already exists in the world. Before you write a word of prompt, read the character card and write down three things: what this person wants in the scene, what they are hiding, and what their hands are doing. The first two decide the blocking; the third decides whether the image looks alive.

For Shōya, the card says: a quiet high-school senior, messy black hair, downcast eyes, a former bully who was then made the scapegoat. In the bridge scene he wants to apologise and to disappear at the same time. That contradiction gives you the staging: he stands too far away, holds the notebook closed, and looks at her shoes rather than her face. None of that is a style word. All of it is visible.

Lesson 2 — Build the frame in three planes

Anime key art reads best when the viewer can separate foreground, subject and background without thinking. Write them in that order. Foreground: the railing, the desk edge, the camera. Subject: the figures, placed at a specific distance from the foreground. Background: the river, the town, the blackboard. If you write the background first, the model tends to paste the figures on top of it like stickers.

The bridge prompt uses the railing as the foreground plane and the river as the receding middle ground. The town’s apartment blocks and power lines sit behind at a lower contrast, which keeps the figures dominant. The cherry tree on the left is not decoration; it stops the frame from feeling like a corridor and gives the light something to pass through.

Lesson 3 — Light is the argument

Do not describe the mood. Describe the light that produces it. “Warm low sun and long soft shadows” gives you late afternoon without naming melancholy. “Cold fluorescent light from above” gives you a classroom that feels like an institution. The model responds to direction and quality far more reliably than to adjectives.

For the bridge, one key and one fill is enough: the low sun from the left, a soft bounce off the pale railing. For the classroom, the window is the key and the green blackboard is the negative space. If you add a second window or a ceiling light, the room stops feeling empty and starts feeling lit for a photograph.

Lesson 4 — Direct the hands

Hands are where anime models fail most visibly, and where this story lives. Give every visible hand a job. Shōya’s hand holds the notebook or rests on the railing. Shōko’s hands are half-raised in a sign, not floating in a generic gesture. Yuzuru’s hand holds a camera. A hand with a job is easier to draw and more meaningful than a hand posed for elegance.

If a hand is central to the scene, say what it is doing and what it is near. “Her hands half-raised at chest height, palms toward him, a hearing aid visible behind her ear” is a block of instructions the model can follow. “She signs apologetically” is not. The model cannot infer grammar from an adverb.

Lesson 5 — Render absence

The hardest image in the set has no people in it. An open notebook and a pair of hearing aids on a desk; cherry petals; empty chairs; a clock. The prompt has to make the room feel recently occupied. The trick is to choose objects that imply a body: a pushed-back chair, an open page, a hearing aid placed rather than dropped. Then let the light do the rest.

GPT Image 2 renders this as a still life with depth: the desk in the foreground, the rows receding, the blackboard holding the back wall. Krea 2 renders it as a colour event: the window becomes a flat poster shape and the objects become bold graphics. Both are valid; only one is a background plate. Decide which you need before you generate.

The prompt template

The repair pass

Never rewrite the whole prompt after a bad render. Change one clause and regenerate. If the composition is wrong, fix the planes. If the faces are wrong, remove a character and simplify the silhouette facts. If the image is beautiful but empty, add an obstruction and a hand task. If the image is busy, remove a light source and one background detail. A repair pass should take one sentence, not a new paragraph.

Style lock for anime consistency

Consistency is a sentence, not a model setting. Write one style lock and reuse it verbatim: “Fine precise anime linework with slightly tapered lines, transparent light local colour, low-contrast value groups, watercolour-like transitions, soft controlled shadows.” That sentence does more for a series than any seed. It tells the model what kind of mark to make, how colour sits on the page, and how much contrast to allow. Change one word and the series drifts.

If you need character consistency across scenes, add a reference image and keep the silhouette facts identical. Do not rely on the model remembering a face from a previous prompt. Treat every generation as a new shoot with the same actor, same costume and same lighting plot.

The two-work comparison method

Checklist before you publish

Questions about this guide