Creative · Guide №0001 · 4 min read
Stop describing your characters. Start showing them.
Keep one character looking like themselves across every scene of a picture book, webcomic, or campaign — free, inside Google AI Studio, with nothing but a Google account: one anchor sheet, one pinned reference image, zero re-roll lottery. The catches, stated up front: the image model is an experimental preview whose name keeps changing, free-tier uploads can be used to improve Google's models, and if you script with an API key you get 15 requests/minute and 1,500/dayas of 2025-06-12.
Numbers wearing an "as of" chip drift. The method doesn't.
The centerpiece — Anchor Builder
Build the sheet. Copy the prompts. Pin the reference.
This is the whole workflow as a tool. Fill in your character once, then copy each output into Google AI Studio at the moment the walkthrough tells you to. Nothing is sent anywhere — the builder runs entirely inside this page.
Why only three anchors? A twelve-trait sheet is a sheet the model half-remembers — every extra immutable is one more thing to drift.
The whole method, one line
More words make a character less consistent. A reference makes it more.
01 — Why references win
A text prompt is a suggestion. An image is evidence.
Image models re-interpret words on every generation, and each adjective you add is another independent chance to drift — which is why the standard advice, "describe your character in more detail," produces characters that fall apart by scene four. A pinned reference short-circuits the interpretation: the model copies what it sees with far higher fidelity than what it reads.
That's why the anchor sheet caps immutable traits at three. One mark, one worn item, one proportion — the three survive close-ups, action, and dusk. Everything else is allowed to serve the scene, which is what makes a character feel alive instead of traced.
And it's why the workflow says correct, never re-roll. A one-line correction ("the beanie stays red — redo this scene only") costs one request and the model keeps the fix for the rest of the session. A re-roll costs the same request and keeps nothing. At 15 requests/minute on the API tieras of 2025-06-12, that difference is your whole day's budget.
What it replaces
Tick what you'd otherwise rent for consistency — then run the swap.
02 — The walkthrough
Five steps, one character
Open the studio, not a chat window
Go to aistudio.google.com and sign in with any Google account — no card, no key, T0. In the model picker, choose the image-capable model; in June 2025 that's the Gemini 2.0 Flash image preview, and the label has changed at least three times this yearas of 2025-06-12, so take whichever model generates images — the method survives the renames.
Before typing anything, open the right-hand settings panel and find System instructions. Skipping that panel is why most people end up re-pasting their character description into every message, where it competes with the scene text and quietly loses.
Write an anchor sheet, not a description
Paste this into System instructions (or use the Anchor Builder above for a filled-in version). Three immutable anchors, maximum — every extra one is another drift surface and another instruction diluting the anchors that matter.
You are illustrating a picture book. Follow this CHARACTER ANCHOR SHEET in every image you generate. CHARACTER: [name] IMMUTABLE ANCHORS — identical in every image, no exceptions: 1. [anchor 1 — a facial mark or feature] 2. [anchor 2 — a worn item that never comes off] 3. [anchor 3 — a body proportion or silhouette note] SIGNATURE: [one oversized/distinct item] PALETTE: [3 colors] STYLE: [art style]; default framing [full-body / waist-up / close-up] RULES: The three anchors never change — not for close-ups, not for action, not for "artistic interpretation." Everything else (pose, expression, setting, weather) serves the scene.
Make the master reference
Send the prompt below and re-roll until you love one — this is the last lottery ticket you buy. Download the keeper at full resolution. That file is now the character; the sheet just describes it. Check the thumbnail: if the anchors don't read at 10% zoom, tighten them and regenerate before building anything else.
Using the anchor sheet above, generate ONE [framing] reference of [name] on a plain neutral background. No text. No props beyond the signature item. Make each anchor unmistakable even at thumbnail size.
Pin, don't describe
Drag the master reference into the chat, then send the scene prompt. The reference rides along in the conversation, so later scenes don't strictly need re-pinning — until drift appears (see What Breaks). Keep scenes short and physical; long scene prose gives the model room to reinterpret the anchors.
[Attach the master reference image to this message, then send:] Same character, same style. Scene: [scene]. Keep the three anchors and the palette exactly. Change only pose, expression, and setting.
Drill, then correct — never re-roll
Run the five-scene consistency drill. If an anchor breaks, correct in one line. A correction costs one request and compounds — the model remembers the fix. A re-roll costs the same request and resets your luck — the model remembers nothing.
Consistency drill — generate these five scenes in order: 1. front-facing close-up 2. side profile, framing matched to #1 3. full-body in action (running, jumping, reaching) 4. the scene from #3, but at dusk 5. peak emotion (laughter, tears) If an anchor breaks, correct in one line — never re-roll: "Anchor fix: [which anchor drifted]. Redo this scene only, everything else unchanged."
03 — Quota math
Will you hit the ceiling?
What breaks
Most guides tell you to fight inconsistency with prompt detail — more adjectives, full wardrobe inventories, "hyper-detailed, 8k." That's backwards. Text prompts are a weak conditioning signal: the model re-interprets your words on every generation, and each descriptor you add is another independent chance to drift. Re-rolling a 40-word description is how an afternoon becomes three usable images. References are the strong signal. Put the detail in the anchor sheet; put the truth in the reference image.
- The model name is perishable. It's an experimental preview; the label changed repeatedly in 2025 and quotas moved with itas of 2025-06-12. Bookmark the method, not the model name — and if a future model gates image output behind a paid tier, the web-app path is the one to re-check first.
- Free-tier privacy. Uploads on the free tier can be used to improve Google's models — nothing you send is private in the NDA sense. Client portraits, other people's children, unreleased product shots: keep them out. You can't un-train a model.
- SynthID. Every output carries Google's invisible watermark. Harmless for books and web; some stock and marketplace pipelines reject watermarked files — that's your mystery rejection letter.
- Drift after long sessions. Past roughly fifteen generations in one chat, earlier references compress and anchors soften. Symptom: the beanie changes color. Fix: re-attach the master reference and restate the sheet in one line — or start a fresh chat per chapter.
- Quota, if you script it. The web app is fair-use unmetered; the API-key path is 15 req/min and 1,500/dayas of 2025-06-12. Careless re-rolls at 15/min burn a morning's budget in minutes.
- Lettering. Don't ask this model class to render text — signs, covers, speech balloons. Generate the art; add type in any editor.
05 — Three things worth keeping
If you forget everything else
The model copies what it sees more faithfully than what it reads. Test it yourself: run one scene from a 40-word description, then the same scene from a pinned reference, and count surviving anchors. The description version loses at least one per scene — across a 20-scene book, that's a different character by page six.
Every extra immutable trait dilutes the ones that matter and adds a drift surface. Pick one mark, one worn item, one proportion. If a trait can't survive a close-up on its own, it isn't an anchor — it's set dressing.
A correction ("the beanie stays red — redo this scene only") costs one request and the model keeps the fix for the rest of the session. A re-roll costs the same one request and keeps nothing. Same price, opposite compounding.
06 — Try it now
Five minutes, two matching scenes
- Open aistudio.google.com — sign in, no card, no key.
- Click "Load example" below — Juno moves into the Anchor Builder.
- Copy the system instruction into AI Studio's right-hand panel.
- Send the master-reference prompt; keep the best output.
- Drag that image back in with the scene prompt — two matching scenes of the same kid before your coffee cools.