How to Make Realistic AI Video With Consistent Characters (Live-Action Style)
Make photorealistic AI video where the same actors appear in every shot and speak their lines: cast, wardrobe, locations, dialogue and real 2026 costs.
To make realistic AI video with consistent characters, cast your actors once as reference sheets, then generate every shot from those sheets instead of describing the people again in each prompt. In Paintbrush you create a project in Realistic Style, make each character a full character sheet (front, side profile, back and a face close-up), write the scene as a script, and the shots are generated in order with the sheets attached. The same faces and clothes carry through every cut, and the characters speak their lines with their mouths in sync. As of November 2026 a character costs 19 credits and a 5-second shot with dialogue costs 90.
Below: why photoreal is harder than cartoon, the workflow step by step, how to make shots look like live action and not like a render, what it costs, and how the main tools compare.
Why is realistic AI video harder to keep consistent?
Three reasons, and they compound.
- People notice real faces. A cartoon character whose nose changes a little between shots still reads as the same character. A realistic face that shifts by the same amount reads as a different actor. We are very good at recognizing faces, so photoreal work has less room for drift.
- Text prompts do not pin a face down. "A woman in her thirties with dark curly hair" describes millions of people. Generate it ten times and you get ten different women. Text-only generators lose consistency fastest in realistic style. See why AI video loses character consistency.
- Photoreal references get refused. The fix for drift is to give the video model reference images of the character. Some hosted video models reject realistic reference images because they might show a real person, even when the face was generated from scratch. This is the most common reason "it worked in cartoon style but fails in realistic" (more below).
Step 1: Create the project in Realistic Style
Pick Realistic Style when you create the project. It tells every image and video model to draw photorealistic live-action photography: real people with natural proportions, real skin and fabric, natural light, and explicitly not a cartoon, 3D render, doll or figurine. Without that instruction, a word like "character" tends to pull image models toward stylized 3D figures.
Choose the aspect ratio for where the video will be watched: 9:16 for TikTok, Reels, Shorts and vertical drama series, 16:9 for YouTube and a film look, or 4:3 and 3:4 in between. Decide now. Shots are composed for the frame, and cropping 16:9 to vertical later cuts off heads and hands.
Step 2: Cast your actors as character sheets
Each character is one reference sheet. You describe the person, and Paintbrush draws a full-body front, then a side profile, a back view and a face close-up from that front. Every shot the character appears in is generated with those images attached, which is what keeps the face the same.
Write the description like a casting note, with specific, visible details:
Daniel, early forties, tall and broad-shouldered, close-cropped grey-flecked beard, a thin scar through his left eyebrow, tired eyes. Navy work jacket over a grey henley, dark jeans, scuffed brown boots.
Specific marks (the scar, the grey in the beard) give the model something to keep. Look at the sheet before you generate any video. If the face is not right, describe the change and regenerate the character. One 19-credit retake now saves every shot that would have used the wrong face.
Can I use a real photo as the character?
Yes. Upload a reference photo when you create the character and the sheet is drawn from it (20 credits), with up to three extra photos of the same person used for the face close-up, 1 credit each. Use photos of yourself or of people who have agreed to it, and never public figures. How to pick the photos is in how to choose reference images.
How do I change a character's clothes between scenes?
Make a variant: another full turnaround of the same person in different clothes, such as "Daniel in a black suit for the funeral". You choose the variant per chapter, or per shot when someone changes during a scene. Do not write the new outfit into the shot description; the attached sheet will win and you will get the old clothes. See how to change a character's outfit.
Step 3: Build the locations
A setting is a reference image of an empty place, drawn wide and at eye level: "a cramped hospital break room at night, fluorescent tubes, a vending machine, rain on the window". It costs 3 credits. Shots in that place are drawn from it, so the room keeps its layout and light across cuts. For realism, name the light source and the time of day: it is the thing that most often gives a generated location away when it jumps between shots. More in consistent backgrounds in AI video.
Step 4: Write the scene as a script
Write dialogue as script lines with a speaker, not as quotes inside a prompt. Each line can carry a delivery tag such as [quietly] or [angry], and silent beats describe action with no words. Give each character a voice from the ElevenLabs Voice Library; for realism, pick a voice whose age and accent match the face.
The whole chapter is performed as one take, so lines answer each other with natural timing. The video model then voices each shot's lines with the speaker's mouth moving in sync. A realistic face with a mouth that does not match the words is what makes AI video feel fake, so this step matters more in live action than in cartoon. See how to make AI characters talk.
If you already have a story or a script, paste it into the creation agent. It plans the cast, the locations and the chapters, and runs the same steps for you.
Step 5: Cut and generate the shots in order
The cut director divides each chapter into shots from its recording, so each shot is as long as its lines (any whole second from 4 to 30). Shots are generated in order, and each is staged from an earlier shot: a continuous shot opens on the last frame of the one before, and a new angle in the same room is first drawn as a still keyframe from that earlier shot (4 credits or more), so positions and light carry over. Watch each shot as it arrives. If one drifts, fix it before the shots after it are staged from it.
How do you make AI video look like live action, not a render?
- Describe shots like a camera operator. "Medium close-up, handheld, shallow focus on Daniel, the window behind him out of focus" gets a more filmed look than "Daniel looks sad".
- Favor medium and close shots. Faces hold best when they are large in the frame. Wide shots with small figures are where likeness slips.
- Keep the action simple. A glance, a step, a hand on a shoulder. Fights, dancing and fast hand work (typing, shuffling cards) are where realistic video breaks first. Cut around them.
- Limit the cast per shot. Two or three people per shot is reliable. A crowd of named characters is not. See multiple characters in one scene.
- Keep light consistent. Same time of day and weather through a chapter. A new time of day is a new chapter.
- Use objects for readable text. A phone message or a sign is drawn as an object first, so it is readable. See readable text on screens.
Why do AI video generators refuse realistic characters?
Video model providers are cautious about deepfakes, so some hosted versions of popular models check reference images for real people and reject anything that looks like one. A generated face can look real enough to fail that check. The result is a tool that animates your cartoon cast but refuses your realistic one, often after you have spent credits on the character images.
Paintbrush runs its default video model, Seedance 2.5, on a host that accepts photoreal reference images, so realistic characters generate like any other style. If a shot is still refused, it is made on Wan 3.0 in the same job instead. The Seedance charge is refunded and the Wan price (6.5 credits per second as of November 2026) is charged. Wan speaks the lines in voices sampled from the recording, so a shot that fell back can sound slightly different from its neighbors. For mature themes, see the Uncensored switch.
What does a realistic AI video cost?
An example, at November 2026 rates: a one-minute scene with two characters in one location, cut into twelve 5-second shots with on-camera dialogue, half of them new angles.
| Item | Credits |
|---|---|
| 2 character sheets (19 each) | 38 |
| 1 setting | 3 |
| 12 shots of 5 seconds (90 each, voices and lip sync included) | 1,080 |
| 6 keyframes for new angles (4 or more each) | 24+ |
| Speech, about 1,000 characters | 8 |
| Direction (cut director and voice director, about 2 chapters) | ~24 |
| Total, before retakes | about 1,180 |
That is about $28 at the smallest pack's rate ($12 for 500 credits). Video is almost all of it, so the way to spend less is fewer seconds, not cheaper characters. Budget one or two retakes per minute. The characters and settings are reused in every later episode, so episode two costs only its shots. Full breakdown in how much AI video costs per minute.
Which tool is best for realistic AI video?
It depends on whether you need one shot or a story.
- Veo 3 is best for a single stunning photoreal shot with sound. It is a flagship model, not a story workflow: keeping the same person across many shots is up to you.
- Kling is best for realistic motion and physical action. Its Elements take reference images of a character for each clip, but you attach them yourself every time, and dialogue usually needs a separate lip-sync pass.
- Higgsfield is best for a reusable photoreal actor in cinematic single shots. Building a dialogue scene from many shots is still your job.
- Runway is best for directing and finishing individual shots by hand. It has no saved cast that follows a character from scene to scene.
- Paintbrush is best for multi-shot stories in realistic style: the same cast in every shot, wardrobe changes, scripted dialogue with voices and lip sync, and shots that follow each other. Its clips are 480p, so for a single hero shot in high resolution, a flagship model will look sharper.
For a broader comparison, see the best AI video generators for consistent characters.
FAQ
Can AI make a realistic video with the same person in every scene?
Yes, if the tool generates every shot from the same reference images of that person. In Paintbrush, each character's sheet is attached to every shot they are in, across chapters and episodes.
Can I put myself in a realistic AI video?
Yes. Make a character from your photo (20 credits, plus 1 for each extra face photo, up to three). Only use photos of people who have agreed to it.
Do realistic characters cost more than cartoon ones?
No. Prices are the same in every style: 19 credits per character, 3 per setting and 18 credits per second of video as of November 2026.
What resolution is the video?
480p. That is fine for phone viewing and social video. It is the main trade-off against flagship single-shot models.
Can the characters talk?
Yes. Write the lines in the chapter's script and give each character a voice. Each shot is generated with its lines voiced and the speaker's lips in sync.
Why does my realistic character look like a 3D render?
Check that the project uses Realistic Style, and describe skin, fabric and light in the character's description. Regenerate the sheet before making any shots: the video can only be as real as the reference.
Cast it once. Shoot every scene.
Create characters, plan scenes, and bring your story to life in one workspace.
Try Paintbrush