If you have ever generated a character you loved and then tried to put that same character in a second scene, you know the problem. The hair changes length. The jacket switches color. The face drifts just enough that it no longer feels like the same person. For a one-off illustration that is fine. For a comic, a mascot, a product storyboard, or a series of social posts, it breaks the whole project.
Image models are not designed to remember. Each generation is a fresh attempt, guided only by what you give it in that moment. So consistency is not something you switch on; it is something you engineer through your inputs. This post walks through a workflow I use, with concrete steps you can adapt to most reference-capable image tools.
Step 1: Write a character sheet before you write a prompt
Before generating anything, write a short plain-text description of the character that you will reuse word for word. Treat it like a config file. A good sheet covers:
- Identity anchors: age range, build, face shape, hairstyle and hair color.
- Signature items: one or two distinctive pieces, like a yellow raincoat or round glasses. Distinctive beats detailed.
- Palette: three to five colors that belong to the character.
- Expression range: the default mood, and what "happy" or "worried" looks like for them.
Keep it short, around 40 to 70 words. Long descriptions tend to get partially ignored, and every extra adjective is another thing that can drift. The point is to have a fixed block of text that never changes between scenes, while only the scene part of the prompt varies.
Step 2: Separate the fixed block from the variable block
Structure every prompt in two parts:
[CHARACTER] A woman in her late 20s, short black bob, round tortoiseshell
glasses, mustard-yellow raincoat, calm half-smile, palette of mustard,
navy and warm grey.
[SCENE] Standing at a rainy tram stop at dusk, soft streetlight from the
left, medium shot, 16:9 framing with space on the right for a caption.
When something goes wrong, you now know where to look. If the character drifted, the fixed block was not strong enough or a scene detail conflicted with it (for example, "bright sunny beach" fighting a raincoat). If the setting is wrong, you only edit the scene block.
Step 3: Generate one hero image and treat it as the source of truth
Run a handful of generations with the character sheet and a neutral scene: plain background, even lighting, a clear front or three-quarter view. Pick the single best result. This is your hero reference.
Don't pick the most dramatic image. Pick the clearest one, where the face, hair, and signature items are easy to read. A reference full of shadows and motion teaches the model less.
Step 4: Use the hero image as a reference for every new scene
This is where reference-capable tools make the biggest difference. In Nano Banana 3, for example, the image generator lets you add a reference image when the selected model supports it, and the site's own guidance is to explain which qualities of the reference should carry into the new picture. That instruction matters more than people expect. Compare:
- Weak: "Use this image."
- Strong: "Keep the character's face, black bob haircut, round glasses and yellow raincoat from the reference. Change the setting to a busy café interior."
Telling the model what to keep and what to change narrows its freedom in exactly the places you care about.
Step 5: Change one thing at a time
Consistency usually breaks when a single prompt asks for too much: a new location, a new pose, a new outfit, new lighting and a new camera angle all at once. Treat each new scene as a small diff from the previous one.
A practical sequence for a short series:
- Hero image, neutral background.
- Same pose, new location.
- New pose, same location.
- New lighting or time of day.
- Close-up for emotion.
After each step, compare the result side by side with the hero image. If a detail drifted, state it explicitly in the next prompt ("the raincoat must stay mustard yellow") rather than rewriting everything.
Step 6: Keep a log of what worked
Most of the value in this workflow comes from iteration, so keep track of it. Save the prompt, the reference you used, the model and settings, and a one-line note on what changed. A workspace with generation history helps here, because you can go back to the attempt that nailed the hairstyle and reuse its exact wording.
A simple spreadsheet works too:
| Scene | Fixed block version | Reference | Result notes |
|---|---|---|---|
| Tram stop | v1 | hero_01 | Glasses lost, re-run with explicit mention |
| Café | v1 | hero_01 | Good, keep |
Step 7: Review at the size you will publish
A character can look consistent in a grid of thumbnails and fall apart at full size. Before you commit an image to a page, inspect it at its real display size. Check:
- Hands, glasses frames and other small shapes.
- Any visible text (signs, labels), which models often garble.
- Edges where the character meets the background.
- Whether your crop leaves room for captions or UI.
Common failure modes and fixes
The face drifts toward a generic look. Your reference is probably too small or too stylized. Use a clearer, larger hero image and mention the face explicitly in the "keep" instruction.
The outfit keeps changing. Signature items should be named the same way every time. "Yellow raincoat" in one prompt and "golden jacket" in the next are two different requests.
The style shifts between images. Lock style words into the fixed block, not the scene block, and avoid stacking unrelated style references.
Scenes feel stiff. Once identity is stable, loosen the scene block: add an action and a camera direction. Stability first, then expressiveness.
Budgeting your iterations
Iteration costs time and, on most hosted tools, credits. It's worth checking how your tool charges before you plan a 20-image series. Some charge per result, and higher resolutions or reference inputs may cost more. Nano Banana 3, for instance, shows the credit cost for the selected settings before you submit, and its pricing page explains how credits are shared across models. Whatever tool you use, plan a few extra generations per scene for corrections.
Wrapping up
Character consistency is less about finding a magic prompt and more about disciplined inputs. You need a fixed character block, one clear hero reference, explicit keep/change instructions, small steps, and a log of what worked. Set that up once and a five-image series stops being a fight with randomness and starts feeling like directing.
Top comments (0)