How to Keep a Character Consistent Across AI Film Shots

Why faces drift between shots

Every generation draws the person again. Nothing carries over from the last clip unless you send it: a reference image, a saved character or element, or a frame to start from. A text description leaves room, and the model fills that room differently each time. The jaw comes back a little wider, the eyes a little closer, the parting on the other side.

Video adds three problems that stills do not have. A clip shows the face from angles your references may not cover, so a head turn forces the model to invent the side it never saw. Wide shots leave only a few pixels of face to match. And light changes across a scene alter how skin tone and features read, even when the shape of the face holds. A character who looks right in a lit close-up can drift in the next wide shot at dusk.

What each video tool gives you for a recurring character

Every major video tool now has a way to carry a character between generations, and they differ in what they take. As of September 24, 2026:

• Kling builds an Element from two to four images, or from one 3 to 8 second video, of a single character. Kling VIDEO 3.0 can bind up to three extra elements alongside start and end frames.
• Google Veo 3.1 has Ingredients to video, which takes "up to three asset images of a single person, character, or product" and makes 8-second clips. It works in the Gemini API and Vertex AI, and in Flow on the Fast and Lite models but not on Quality.
• ByteDance's Seedance 2.0 accepts up to nine images, three videos and three audio files as references, and its successor Seedance 2.5 takes up to 30 images for reference-to-video.
• Runway's References feature, in its image tool, takes up to three reference images per generation. Its Gen-4.5 video model works from text or a first frame, so the character reaches the clip through the still you animate.
• Midjourney V8.2 takes up to four reference images, replacing its older Omni Reference and Character Reference.
• Higgsfield's Soul ID asks for at least 20 photos of one person.
• OpenAI removed Sora 2 and its Videos API from the API on September 24, 2026, with no replacement named.

The vendors disagree about what a good reference looks like, and that changes how you use a character sheet. Kling recommends "a front-facing image for higher consistency." Luma's reference guide says to "Show only one angle per image" and "Do not combine multiple angles into one sheet." ByteDance's Seedance 2.0 guide says "using a headshot + full-body photo is sufficient. Using multi-view character images is not recommended," while Seedance 2.5 accepts multi-view images. Google's Flow asks for references "on a plain or segmented background," and Runway's guide asks for "Natural, even lighting" and a "Neutral subject expression."

So keep the character sheet as the master record and cut it into what each tool wants to receive: a clean front-facing headshot, a full-body shot, and single-angle crops, each on a plain background.

Cast in stills with a character sheet

Before any scene, make one image of the character from several angles: front, three-quarter, side and back at full length, plus close-ups of the face, on a plain background in soft, even light, wearing the exact costume. End the prompt with "no words on the image", because sheets tend to come back with labels printed across them. Approve it, save it with a name and a version number, and treat it as the casting decision it is.

Make one sheet per costume. If the character changes jackets in act two, that is a second sheet, built from the first, with the new jacket.

The agency 100 Creatives counted what this step takes on a 2026 surf apparel job, one model across studio, cliff and ocean scenes. The character sheet took 20 attempts. The first attempt had the longest prompt and four reference images; later attempts went up to eight references. The kept sheet, attempt 20, used five references and a four-line prompt. Fewer, better-chosen references and a short prompt beat the long versions.

Start every shot from the sheet

The sheet only works if every shot draws on it. Image-to-video gives you the most control. Make each shot's first frame as a still from the sheet, check the face at full size, then animate that frame with a prompt that describes motion only. A 1K still on Google's Nano Banana 2 costs $0.067; an 8-second Veo 3.1 clip at 1080p costs $3.20 at Google's published $0.40 a second. A face you reject at the still stage never reaches the edit. Our guide to clean first frames covers the frame itself.

Then lock whatever is not changing. Give the character a name and one description sentence covering age, build, skin tone, hair and face shape, and paste that sentence into every prompt unchanged. Change only what the shot needs: framing, pose, action. When a scene continues across clips, go back to the sheet for the next shot's first frame instead of extending from the last frame of the previous clip, because small changes add up when each clip starts from the one before.

Plan coverage around what the model can hold. Close-ups and medium shots give it more face to match than wides do. A turn that shows the back of the head needs a back view on the sheet. Cutting on action is safer than asking one clip to carry a full turn.

Real faces, likeness and contest rules

If your lead is a real person, the rules come from three directions. Several tools restrict real faces. Google's Veo guidelines say it "Rejects requests to generate a photorealistic representation of a prominent person," and ByteDance's Seedance documentation says the models "do not support direct uploads of reference images or videos containing real human faces," offering preset digital characters and authorized portrait assets instead.

The contests ask for consent. The Formula E Creators Challenge requires written consent for "the name, image, voice or likeness (including any AI-generated likeness) of a real person." Curious Refuge's 2026 horror rules say "You may not use the likeness of existing actors or people without their consent" and bar the likeness of any famous person currently living. And festivals point you to the law: Runway's AI Festival terms require compliance with "any applicable digital replicas/deepfakes and data protection laws."

The simple route for a contest film is a fictional character, built from scratch and approved as a sheet. If you cast a real actor, get their written consent for AI use, name the project in it, and keep it with your files.

Budget the tries before you spend on video

Holding one person steady across scenes costs tries. Across four 2026 campaigns at 100 Creatives, 361 renders produced 45 usable images, about 8 per keeper. Surf apparel on one model across studio, cliff and ocean scenes took the most, 13.6 renders per keeper (163 for 12), while gummy supplement ads took 3.2 (38 for 12).

Use those numbers to plan. As an example, if your lead appears in 30 shots and you make each first frame as a still, 8 to 14 tries per usable frame means 240 to 420 stills before a single clip renders. Stills cost cents each, and every face you fix at the still stage is a video render you do not pay for twice. Contest deadlines make this sharper. Curious Refuge's AI Horror Film Competition, which closes October 9, 2026, caps films at three minutes and requires AI in every shot, so every shot is a chance for drift.

Where Overs fits: the character sheet, made and approved first

Overs, an AI tool for product and campaign photos, is built around the step this page keeps coming back to. For any product that is worn or held, it makes the character sheet first: one model, several angles and close-ups, wearing the exact product, on a plain background. You approve the sheet, and every later photo with a person uses that same model. Approved sheets go into a talent library, so the face from one campaign can come back for the next.

Each photo gets only the reference pictures it needs, in order, with the sheet first for any shot with a person, and if a reference the plan calls for does not exist, Overs skips it instead of guessing with the wrong picture. You see the shot plan and an estimated cost before anything renders, and nothing is charged until you say yes.

Overs makes stills, which is the right split for a film. Approve the character in stills, then animate the frames in your video tool, starting from the text motion prompt Overs writes for each photo for a model such as Seedance or Veo. The method comes from 100 Creatives, the agency Abhi Chawla founded, which used it to make hundreds of photos a week for hundreds of brands, and the surf job above is one of its campaigns. At its average of about 8 tries per usable photo, published prices of $0.034 to $0.211 a render come to $0.27 to $1.69 in AI fees per keeper. Abhi's own businesses make their images with Overs today, and he credits it for their growth. Overs' guide to keeping the same model in every photo has the kept sheet prompt and a drift fix table.

Overs is a sister product to AI Film Contests, from the same founder. The free plan gives you 40 photos a month. You connect your own OpenRouter key and pay the model a few cents a photo, and Overs adds no markup. Try Overs free at www.overs.studio and cast your lead before your next deadline.