Pencil0
For video, across shots and episodes

Consistent Character AI for Video

Plenty of tools keep a face the same across still images. Keeping it the same across dozens of video shots, in different rooms and different lighting, is harder. Pencil0 builds the whole workflow around that: cast once, then reuse the cast in every shot.

In private beta · leave your email, be first in when we open

Email only, no spam · Already have an invite code? Sign up

SC 01How it works

Cast once, then reuse the cast in every shot

Consistency here is a process, not a switch. The mechanism has three parts, and the last one is you.

  1. 01
    STEP 01

    Cast the characters

    Characters, scenes and props are extracted from your script. Each character gets a reference image, either generated from its description or uploaded by you. You approve it before any video exists.

  2. 02
    STEP 02

    Reference them in every shot

    When a shot is generated, the reference images of the characters in that shot, plus the scene reference, are sent to the video model, with instructions about what each image controls: a character image locks that character's face, hair, build and clothing.

  3. 03
    STEP 03

    Review, then regenerate what drifted

    Watch each shot. If a face wandered, regenerate that one shot. You do not have to redo the episode, and the cast reference stays exactly as you approved it.

SC 02What carries the character

More than one thing keeps a character stable

A single reference image helps, but a cast that survives a whole story needs several mechanisms working together.

  • One approved reference per character

    Every shot cites the same reference asset for a character, whether you uploaded it or generated it. The face in shot 40 is conditioned on the same image as the face in shot 1.

  • Scene references too

    Locations get reference images as well, so a room keeps its layout and lighting between shots. This stops a stable character from standing in a different room each time.

  • Storyboards that carry no faces

    The black-and-white storyboard frame shows only position, posture and framing, using blank placeholder figures. A face on the sketch can only come from the character reference, so the sketch cannot introduce a different face of its own.

  • Continuity from an earlier shot

    For models that accept it, you can pick an earlier accepted shot from the same episode as a video reference, which helps a following shot match the one before it.

  • Costume changes without changing the person

    A character can have extra looks, such as a different outfit. A look is generated with the character's default image attached, so the identity stays and only the described appearance changes.

  • A story bible across episodes

    In a series, the project records each character's details and their state in every episode, and uses it when new episodes are written. That keeps the script consistent, which gives the images less to contradict.

SC 03Craft notes

Why characters drift in AI video, and what actually helps

It is worth knowing why it happens, because the fixes follow from the cause.

NOTE 01

How to prepare character reference images before generating video

Start with one approved image of one character: a readable face, visible hair and an outfit whose colors you can describe. Record three or four identity details, such as Mara's short dark hair, green coat and brass whistle. Keep that master image when reviewing new angles; a rejected variation should not become the reference for the next attempt.

Separate identity from shot instructions. Keep the face, hair and costume stable while changing the framing or action. Test a medium shot and a reaction close-up before expanding to an episode. Compare the same facial features and costume details in both; review motion separately after video generation.

In the asset editor, Upload final saves an accepted image directly. Reference + AI generates a new image using the current image model, so its output needs another review. The character-reference preparation guide provides copyable prompts, a three-shot check and the distinction between an asset reference, a storyboard and a first frame. It uses FLUX 3 Klein externally; Pencil0 does not currently integrate that model.

NOTE 02

The model generates each shot from scratch

A video model does not remember the last shot unless you hand it something to go on. Each clip is a new generation, guided by whatever you give it: the text prompt and the reference images. If the only identity information is a written description such as "a woman in her thirties with short dark hair", every shot will interpret it a little differently, and the differences accumulate.

That is the first thing a reference image fixes. It replaces an ambiguous description with a concrete face, and it is why this workflow asks you to settle the cast before generating video.

NOTE 03

The prompt can fight the image

Drift is also caused by shots that ask for something the reference cannot support: an extreme angle, a heavy change of light, a character turning away for most of the clip. The model fills the gaps, and the filling is where faces change.

So write shots that the reference supports. Keep the named characters in a shot few, avoid cramming a long scene into one shot, and treat a close-up with a character you have never shown from that angle as a risk.

NOTE 04

Reference capacity differs by model

Models accept different numbers of reference images, which limits how many named characters can be pinned in a shot. The specs below list what each model in Pencil0 accepts. A shot with several named characters and a scene reference needs a model that takes several images.

NOTE 05

Accept that some shots will need another try

No current video model guarantees an identical face on every generation. Pencil0 reduces drift, it does not abolish it. The sensible workflow is the one built in here: approve the cast, generate, review, and regenerate the shots that missed. Because each shot is generated separately, a miss costs you one shot, not the episode.

For the wider workflow, see the AI series generator for multi-episode shows and the AI movie maker for single films.

NOTE 06

Review the voice as well as the face

Visual and vocal continuity need separate checks. Keep a cast sheet for voices as well as reference images, and compare each accepted take with the same character in adjacent scenes. The Drama 3 performance guide includes a cast sheet and listening checklist for an external voice workflow; it does not describe a Drama 3 integration in Pencil0.

SC 04Specs and limits

How many references each model accepts

These are the reference limits Pencil0 applies per shot for each video model.

MINIMAXMiniMax Hailuo H3
Up to 9 reference images and 1 reference video. This is the default model.
SEEDANCESeedance 2 Fast, Seedance 2, Seedance 2.5
Up to 30 reference images and 1 reference video.
REF-VIDEOReference video
At most one earlier accepted shot from the same episode.
REF-IMAGEReference images
Generated from the character description, or uploaded by you.
COSTCost
Credit-based, priced by shot length and resolution. There is no free tier; new accounts can buy a one-time trial pack on their first purchase, sized to finish one complete short film. See pricing.
ACCESSAccess
Currently an invite-only beta. Join the waitlist above, or sign up if you already have an invite code.

Model limits can change as providers update them, so treat the pricing page and the studio as the source of truth.

SC 05Questions

Consistent character AI: frequently asked questions

Q.01

How do I keep the same character in AI video?

Give the character a reference image and use that same image for every shot they appear in. In Pencil0 you cast the characters first, approve their reference images, and the references are sent to the video model with each shot. You then review the results and regenerate any shot where the face drifted.

Q.02

Can I upload my own character reference?

Yes. A character reference can be generated from its description or uploaded by you, and once it is in place every shot cites that same asset.

Q.03

Does it work across episodes?

Yes, in the sense that the same character reference is reused in every episode, and the story bible keeps each character's details consistent as new episodes are written. As in a single episode, an individual shot can still drift, in which case you regenerate it.

Q.04

Which video model should I use for consistency?

There is no single answer, and we do not publish a ranking. What we can state is capacity: MiniMax Hailuo H3 accepts up to 9 reference images, and the Seedance models up to 30. For shots with several named characters, pick a model that accepts several images, and test a short shot before committing to a long episode.

Q.05

What if a shot drifts?

Regenerate that shot. The reference images are not changed by regenerating, and the other shots are not affected. If the same shot keeps drifting, look at what it asks for: an unusual angle, a heavy lighting change or many named characters at once often explains it, and simplifying the shot usually helps.

Q.06

Can a character change outfits?

Yes. A character can have additional looks. A look is generated with the default image attached so that the identity is kept and only the described appearance changes.

Q.07

Should I upload a character sheet or a single reference image?

A multi-view sheet is useful for human comparison, but a clean single-character image makes the intended reference easier to read. Prepare separate views yourself when needed; this guide does not establish automatic splitting of uploaded character sheets. Use the same approved master when checking new shots.

REEL END — 00:00:00:00

Cast yourcharacters once

Describe your story and approve the cast before a single shot is generated.

Join waitlist

Free to join · we'll email you the day we open