AI character consistency: how to keep the same character across scenes
By Arsène Huot, founder of TubeTube · Published · Updated
Why does my AI character's face and costume change between shots?
Because each shot is generated on its own: unless the earlier images are sent again, the model never sees them, so it re-invents the face and the outfit from a prompt that describes them loosely. Give every shot the same things to copy (one locked reference image, an identical description of the outfit down to the colours, and the earlier scenes fed back in as context) and the face and costume stay far more stable from shot to shot.
What is character (identity) drift?
Character drift is when a character's identity (face, hair, outfit, proportions, art style) changes from one scene to the next, so a multi-scene video looks like it stars several different people. For a single image it doesn't matter, but for a 2-minute story across 20+ scenes it's the difference between something publishable and something that looks broken. Drift and the rest of the AI video vocabulary are defined in the AI video glossary.
Why do AI models lose character consistency?
Three causes stack up. First, generation is stateless: the image and video APIs TubeTube calls keep no memory between requests, so each scene only knows what is sent with it. Second, there's prompt ambiguity, because “a young girl in a forest” describes millions of different girls. Third, plain sampling randomness: even with the same prompt, the model can land on a different face. None of that is specific to one model, which is why it's more a workflow problem than a model problem.
How do you keep a character consistent across scenes?
Five methods, roughly in order of impact:
1. Lock a reference image (or character sheet)
Generate one definitive image of the character you love, ideally a clean front view, and reuse it as a reference for every scene. A small 'character sheet' (front, side, key outfit) gives the model more to anchor to than a single angle.
2. Generate sequentially, with visual memory
Instead of generating each scene independently, feed the previously generated scene images back in as context. The model then matches what already exists rather than inventing a fresh character every time. In our pipeline this is the single biggest lever against drift.
3. Pin a small set of reference images
Most current image models accept reference images. Pinning a few (the character, the world, a key prop) keeps identity stable across a long video without over-constraining the composition. TubeTube accepts up to 5.
4. Write identity into the prompt, and keep it fixed
Describe the character with specific, repeatable details (age, hair, exact outfit, art style) and copy that block verbatim into every scene prompt. Vague prompts ('a girl') drift, locked descriptions hold far better.
5. Hold the style steady
Keep the same visual style across all scenes. Switching style mid-video is a common, avoidable cause of a character looking like a different person.
How does TubeTube keep characters consistent automatically?
TubeTube bakes that workflow into its image step so you don't do it by hand. Here is what the pipeline actually does, scene by scene:
- Sequential generation with visual memory. Scene images are generated in order, never in parallel. Each new scene is sent with up to the five most recent scene images as context (trimmed to the number of references each image model accepts), so it matches what already exists.
- Up to 5 pinned reference images. You can pin up to 5 images of your own, taken from scenes of earlier videos or uploaded. They sit at the head of the context for every scene, including the very first one, which is how the same character carries from one episode to the next. They aren't billed as scenes. Models that accept fewer images keep the first reference and the latest scenes (Flux 2 Pro takes 2 images in total, Seedream 5 Lite 3).
- A written consistency rule on every reference edit. Whenever references are sent, the prompt asks the model to keep every returning character identical (face, skin tone, hair, age, build, outfit) and to invent a new face only for a character the storyboard explicitly tags as new, so a newcomer doesn't get cloned from an existing face.
- Required appearance, re-injected every time. In Narration + Dialogue mode you describe each named character's look once. That exact text is appended to every scene where the character's name appears, so the model never has to guess the outfit again.
- Character anchors. Also in Narration + Dialogue mode, the first image in which a named character appears becomes that character's visual anchor for all later scenes, even when the character disappears for a while. With the Gemini image models, a solo portrait of each described character is generated first and used as that character's identity anchor instead.
- Spatial continuity in the same location. When two consecutive scenes open on the same place, the new image is told it is the next shot of the same sequence: same architecture, furniture and lighting, and each character stays in the same spot and on the same side of the frame. No teleporting between shots.
- Identity fixed before motion. Every video engine animates the finished scene image, so the character is set at the image step and the clip starts from it.
When a scene image is blocked or fails, the pipeline retries with softened prompts, can switch to another image model, and as a last resort reuses the previous scene image so the edit stays in sync. Every step is logged in the video's transparency report. See it on real videos in the community gallery or in story to video. It's also what keeps the same Moses on screen from the first scene to the last in an AI Bible video.
Which image models can you use, and what do they cost per scene?
You pick the image model per video. The raw API cost of each GPT Image 2.5 tier is measured on the GPT Image 2.5 pricing page. Prices below are TubeTube's credit cost per scene image, read from its pricing table on 23 September 2026 and updated for the 24 September model change (one credit is about $0.011 at the base rate). At the wizard's default 5-second scene length, a 2-3 minute video has 24 to 36 scenes, so at 7 credits an image the default model adds roughly 168 to 252 credits. For the whole video: A 2-3 minute video costs a median 807 credits, about $9 at the base rate of $0.011 per credit; half of them cost between 685 and 1,129 credits ($7.50-12.50), measured on 246 videos in August-September 2026. The exact quote is shown before launch and unused credits are refunded.
| Image model | Credits per scene image | Note |
|---|---|---|
| GPT Image 2.5 · Low | 7 | Default model for new videos, and the cheapest |
| GPT Image 2.5 · Medium | 9 | |
| GPT Image 2.5 · High | 15 | Suggested in the app for photoreal styles |
| GPT Image 2.5 · Max | 50 | Suggested in the app for photoreal styles |
| Gemini 3.1 Flash | 13 | Gemini family: a solo portrait anchor per described character |
| Gemini 3 Pro | 25 | Gemini family: a solo portrait anchor per described character |
| Nano Banana 2.1 | 14 | Gemini family: a solo portrait anchor per described character |
| Flux 2 Pro | 8 | |
| Seedream 5 Lite | 8 | Edits a reference image, so it needs one |
The consistency mechanisms above apply whichever model you choose. The only model-specific one is the solo portrait anchor, which runs with the three Gemini-family models (Gemini 3.1 Flash, Gemini 3 Pro and Nano Banana 2.1). For a comparison of AI video tools on long videos, see the AI long-form video tool comparison.
Real examples: pixar-style 3d made with TubeTube
Finished videos from the community gallery, each with its full recipe (style, engine, voice, lyrics or script) that you can remix with your own words.
Frequently asked questions
What is character drift in AI video?
Character (or identity) drift is when a generated character's face, hair, clothing, proportions or art style change from one scene to the next, so a multi-scene video looks like it stars several different people. It's one of the most visible giveaways of low-effort AI video.
Why do AI characters change between scenes?
Because each generation request only sees what you send with it. The image and video APIs TubeTube calls keep no memory between calls, so unless the previous images and a fixed description are passed in again, the model re-invents the character from the prompt alone, and prompt ambiguity plus sampling randomness do the rest.
My AI generations keep changing my character's costume and face between shots. What fixes that?
Costume and face drift usually come from the same two gaps. The shots are generated independently, so nothing carries the previous look forward, and the outfit is described loosely enough that the model re-invents it every time. Close both. Lock one reference image of the character and reuse it in every shot, then paste an identical identity block (age, hair, the exact outfit down to the colour, the art style) into every scene prompt instead of rewriting it each time. Generating scenes in order, with the earlier images fed back in as context, is what makes it hold. TubeTube does both automatically: it re-sends the recent scene images with each new scene and, in Narration + Dialogue mode, re-injects the appearance you wrote for each named character, word for word, into every scene where that character appears.
How do you get the same character in different scenes with AI?
Give the model something to copy so it never starts from zero. Lock one reference image, or a small character sheet with a couple of angles, and reuse it in every scene. Keep the identity description word for word identical, and hold one visual style across the whole video. Then generate scene by scene rather than all at once, feeding the earlier images back in as context.
How many reference images do you need for a consistent character?
One strong reference is usually enough to start. A few more (the character from a couple of angles, plus the world and any key prop) are more robust across a long video. TubeTube lets you pin up to 5 of your own reference images, taken from scenes of your earlier videos or uploaded, and puts them at the head of the context for every scene, including the first one.
Can Kling or Veo keep characters consistent on their own?
In TubeTube they don't have to. Every video engine there (Kling, Veo, Hailuo, Seedance, Gemini Omni Flash) animates a scene image that has already been generated, so the character's identity is fixed at the image step, before any motion is rendered. Consistency is therefore decided by how the images are generated: in order, with earlier scenes, pinned references and the written character description as context.
Does TubeTube keep characters consistent automatically?
Yes. Scene images are generated one after another, each with up to the five most recent scene images as context (trimmed to what each image model accepts) plus up to 5 reference images you pin. In Narration + Dialogue mode, the appearance you write for each named character is re-injected verbatim into every scene where that character appears, the first image of a character becomes its visual anchor for later scenes, and the Gemini image models also get a solo portrait of each described character as its identity anchor. When two consecutive scenes happen in the same place, a spatial-continuity rule keeps the room layout and each character's position. Image models cost from 7 to 50 credits per scene image; the default, GPT Image 2.5 Low, costs 7.