Consistent characters

AI Video Generator With Consistent Characters

Updated November 18, 2026

Quick answer: character drift is what stops most AI video from becoming a series: by episode 5 the lead has a different jaw and the audience stops believing it is the same show. SceneMixer attacks it with four mechanisms — a cast generated once as reference sheets, face and body crops attached to every segment that character appears in, references linked by identifier rather than by name, and an explicit rule against naming people who are not in frame. Deliberate changes get their own reference art instead of a prompt.
Still from “The Glass Floor”, an AI short drama generated end-to-end by SceneMixer (5 characters, 1 episode)
SampleThe Glass FloorDigital HD1 episodeProduced end-to-end by SceneMixer from the scriptWatch the episodes
SceneMixer at a glanceWhat you get
Video priceFrom $0.06 a second (10 credits) on the default tier, up to $0.47 a second on the top tier; every price is shown before you press the button
Clip lengthUp to 30 seconds per generated segment; an episode is composited from its segments at 1 credit a second
One episode, worked outA 60-second episode on the default tier is about 660 credits (video plus compositing), about $4.11 at the lowest credit price; a 10-episode series of 90-second episodes from a 30,000-character novel is about 10,026 credits, about $62
DialogueSpoken in the video itself, one voice per character, in any of 15 languages
Formats9:16 vertical or 16:9 landscape; 480P to 1080P, plus a 2K tier
Free to start50 credits on signup, 15 free setup images, a free storyboard for the first episode and a free 5-second video preview
Who uses it2,900+ creators have generated with SceneMixer (accounts with at least one generation, counted November 2026)
InputA script, a novel or a one-line idea; .txt, .docx and .md uploads

Why characters drift in the first place

A text-to-video model renders each generation independently. Describe a woman in her thirties with a scar through one eyebrow in shot 1 and again in shot 30, and you get two women who match the description and not each other — adjectives underdetermine a face. Every fix below is the same idea in a different place: stop describing the character, and hand the model a picture of them.

Four locks

What each mechanism does
Reference sheetEach character is generated once as a full-body sheet in 9:16 on a plain background under directional key light — a casting photo, not a mood piece, because whatever light is baked into the sheet gets inherited by every shot that uses it.
Attached cropsFace and body crops from that sheet are sent with every segment the character appears in. The model matches a picture instead of interpreting a description.
Links by idStoryboards point at assets by identifier, not by name. Two characters with the same name, or a rename halfway through a project, no longer break the chain silently.
Off-screen ruleA character who is only being looked at, spoken to or reached for — and is not in frame — is written as a direction, never named. Naming an absent character is what makes a model paste their face into a single-person shot.

A fifth, smaller one: SceneMixer gives each person in a shot one or two distinguishing facial notes pulled from their own description, which takes the edge off the classic failure where two characters in the same uniform trade faces mid-shot.

When the character is supposed to change

Consistency cannot mean "never changes" — a character who is set on fire in episode 6 has to look burned in episode 7. The dividing line is drawn by how a video model treats a reference image: it copies the face, the clothing, the form and the structure almost literally, and it applies light, weather and time of day as effects layered on top.

  • Give it a state variant when what changes is appearance, wardrobe, form or structure: a full change of outfit, a transformation, burns, a collapsed room, a blade out of its sheath, a snapped object. SceneMixer gives the variant its own reference image, generated from the base art.
  • Write it in the shot text when what changes is light, weather, time of day, position, what someone is holding, or what has landed on them. Rain does not need a second version of the street.

Getting this backwards is expensive in both directions: a variant for "holding the cup" is an image you paid for and will never see, and a prompt for "after the fire" is a request the model will politely ignore in favour of the clean room in its reference.

What is still not solved

  • Long coats swallow the face. A full-body 9:16 sheet of a character in a long coat or robe spends most of its pixels on fabric, and the head ends up small enough that fine identity cues are lost. It is a framing limit, not a prompting one.
  • Two characters dressed identically in one shot can still trade faces. Distinguishing notes help; matching uniforms remain the hardest case.
  • Crowds. SceneMixer's consistency guarantees cover the cast in your asset library. Background extras are generated fresh each time and should be treated as scenery.

The layered explanation, with what each layer can and cannot catch, is in the character-consistency guide.

What it costs

Reference sheets are their own line item, separate from script analysis (4 credits per 3,333 characters of English text): each account gets 15 reference images free, and extra or regenerated images after that are 7 credits each on the default image model. State variants carry their own small free allowance on a first project and are billed as images after that. Video itself starts at 10 credits per second of output on SceneMixer (≈ $0.06/s at the lowest credit price) whether or not references are attached — consistency is not a surcharge. Details on the pricing page.

Prove it on your own cast, for nothing

Consistency is the one claim in this category you should never take on trust, and you do not have to. A new account carries 50 credits, 15 free reference images across characters, scenes and props, a free episode-1 storyboard on your first project, and one free preview of up to 5 seconds that renders the opening shot of your own episode against your own sheets. That is enough for a real test:

  1. Paste a scene in which the same two people appear several times, minutes apart.
  2. Let the analysis build the cast, then read the character descriptions and fix anything wrong before art is drawn — the sheets are generated from that text.
  3. Generate the sheets. A reference image takes about 16 seconds at the median, and 9 in 10 finish inside 27 seconds, so a five-person cast is a couple of minutes.
  4. Spend the free preview on a shot where a character is doing something, not standing still, and compare the face against the sheet.

For calibration: the median project on this platform has five named characters, and nine in ten have fewer than ten. If your story needs thirty, the same mechanism still applies — the cost is the sheets, not the consistency.

Every episode, in fifteen languages

The dialogue language is set once in Step 1 and the whole run is performed in it, with the target-region setting moving names, faces and streets to that market at the same time. Each sample below was produced that way, one per language.

Languages

Native dialogue in 15 languages

Every series below was produced by the same SceneMixer pipeline, with the cast speaking that language natively: the interface, the working documents and the spoken lines are all in one language, and nothing is dubbed. Open one to hear it.

See all 15 languages

In SceneMixer vs. stitching tools together

Most AI drama workflows are five or six tools in a row. This is where each step happens here, and what it replaces.

StepIn SceneMixerDoing it by hand with separate tools
Script and episode planPaste a script or a novel, or type one line: it is read into a cast, locations, props and an episode outline you can editA chat model writes the script; you split it into episodes and carry the text from tool to tool
Character faces and placesEvery character, location and prop is drawn once as a reference sheet and attached to every shot it appears inAn image model, one sheet at a time; you upload the right reference to every clip yourself
ShotsA storyboard per episode with numbered shots and seconds; video is generated segment by segment in 9:16 or 16:9 against the same sheetsA video model, one clip per prompt, describing the same people again each time
Voices and dialogueLines are performed in the video itself, one voice per character, in any of 15 languagesA text-to-speech tool for each line, then a lip-sync pass on every clip
Fixing one shotEdit that shot's text and regenerate the segment alone; the rest of the episode is untouchedRegenerate the clip, then re-cut the timeline
FinishingSegments are composited into one episode file with its score; members download without the AI badgeA video editor to join the clips, add music and captions, then export
PublishingYou upload the finished episode to TikTok, YouTube, Instagram or anywhere else; nothing is posted for youUpload by hand, the same as here

Frequently asked questions

How consistent is it, really?

The same face across an entire series is the design target and what SceneMixer's reference-sheet architecture delivers for the cast in your asset library. It is not a guarantee on every frame: identically dressed characters sharing a shot, and characters whose sheet is mostly coat, remain the hard cases. Everything else — episode count, elapsed time, how many shots a character appears in — does not degrade it, because every shot is matched against the same picture.

Can I upload my own character image instead of generating one?

Yes. Any reference sheet can be replaced with an image you upload, and SceneMixer treats it exactly like a generated one. That is the route for a character you have already designed elsewhere.

Does it work with a real person's face?

Technically the upload path does not care what the image is, but publishing AI video of a real person's likeness is a legal problem in most markets and our compliance guide treats it as one. We do not provide a face-swap tool, and we do not recommend it.

What happens if I rename a character mid-project?

Nothing breaks. SceneMixer's storyboards reference assets by identifier, and the rename is propagated through the stored text at the same time. Before September 2026 references were matched by name, and that is exactly the failure this change removed.

Does keeping a character consistent cost extra per shot?

No. On SceneMixer, attaching reference art to a segment does not change the per-second rate — you pay once for each reference sheet, then nothing further no matter how many shots use it. Consistency is an architecture here, not an upsell.

Do I have to re-describe the character in every shot?

No, and you should not. The shot text names the character; the reference art carries what they look like. Repeating a physical description in the shot text pulls the model back towards the words and away from the picture, which is precisely the drift this design removes.

Can one character have two looks in the same episode?

Yes — that is what state variants are for. Each look is its own reference image, and SceneMixer's storyboard decides per segment which one applies, down to the individual segment rather than the whole episode.

Does it work for animated AI characters?

Yes. A reference sheet holds a cartoon, anime or 3D character the same way it holds a realistic one; on SceneMixer the look is chosen once for the project and every sheet is drawn in it.

Same face in episode 1 and episode 40

Generate the cast once; every shot after that matches a picture, not a description.

Try SceneMixer free

Credit packages from $1.49 · Pro $7.99/mo · No credit card to start