What exactly is a locked reference frame?

A locked reference frame is one generated image of a character - face, body, and a specific costume - chosen as the single source of truth for every later shot. Instead of prompting "a woman in a red dress," production teams prompt "this exact woman, from this reference, in this outfit, from a new angle." Every shot re-anchors to that same file instead of asking the model to imagine the character again from scratch.

Why can't a single prompt keep a character consistent across a whole video?

Each generation call is an independent probabilistic draw with no memory of the last one. Ask a model for the same character twice and it returns two different faces, because nothing in the system tells it "match the previous output" unless a reference image is explicitly fed back in. A 3-minute video needs 20-50 shots to look like the same person, which a single prompt cannot enforce.

How much does solving character consistency add to an AI video budget or timeline?

On a mid-tier production, holding one character steady typically costs 3-8 regenerations per shot, and a multi-location narrative video can run 200+ generations in total once every shot is counted. That supervision work - not the prompt-writing - is what turns a same-day social clip into a production with a 5-14 business day timeline.

AI69 / Field Notes 2026-09-02 7 min read

Character Consistency in AI Video: The Hardest Technical Problem, Explained

A character's face can change 3 times in 3 shots with the same prompt. Here is why AI video generation drifts, and the locked-reference process that stops it.

Character consistency - keeping the same face, body, and outfit recognizable across every shot - is the single hardest unsolved problem in AI video generation today. Every generation is an independent probabilistic draw with zero memory of the last one, so a character prompted the same way twice can come back looking like two different people. Left unfixed, a 30-shot video can show 5-10 distinct faces standing in for one artist.

Why does a character's face change between AI-generated shots?

A text-to-video or image-to-video model has no persistent state between calls. The prompt "a woman with dark hair in a leather jacket" describes a category, not an individual - the model samples a plausible match from that category fresh, every single time. Regenerate the same prompt ten times and ten different faces come back, all technically correct, none of them the same person. That randomness is what makes these models good at variety and bad at continuity, and continuity is exactly what a video needs.

What is a locked reference frame, and why does it fix drift?

A locked reference frame is one specific generated image - a face, a body, a costume - chosen and frozen as the single source of truth for a project. Every subsequent shot feeds that image back into the generator as a conditioning input, so the instruction changes from "generate a woman" to "generate this exact woman, from a new angle, in this new setting." On AI69's five-world "Alive Tonight" video for AGA BORYN, the reference sheet locked the artist's face and body plus a separate reference for each of five costume variations, before a single narrative shot was generated.

Locking the reference does not eliminate drift on its own - it gives a human reviewer something concrete to check each output against. Without that anchor image, "does this look right" is a subjective guess. With it, it's a side-by-side comparison, which is what makes the next step - regeneration and selection - actually possible to run at scale.

How many regenerations does it take to hold one face steady?

On a mid-tier AI69 production, a single shot typically needs 3-8 regenerations before one clears review against the locked reference. The number climbs with camera angle: a straight-on medium shot close to the reference angle clears faster, while a low angle, extreme close-up, or a shot with the character mid-motion needs more attempts because the model is extrapolating further from what it was anchored on. A five-world narrative video can run 200+ total generations once every shot in the piece is counted - the bulk of that count is consistency work, not creative exploration.

Does a bigger, more expensive model fix consistency on its own?

No single model solves this by scaling up. Different generators handle character consistency differently: some hold a face well across small camera moves but drift hard on a full angle change, others hold identity but lose fine costume detail like fabric texture or pattern placement. Production pipelines that need reliability combine models - AI69's Cinematic and Signature packages layer Seedance 2.0, Veo 3, Omni, and Kling 2.5 specifically because no single stack wins on every shot type. Picking the right model per shot, not per project, is part of what a character bible is built to inform.

What breaks first: face, body, or costume?

Face identity is the most protected element and the first thing a reference-locked pipeline checks, because audiences clock a wrong face instantly. Costume detail is the most fragile - a specific pattern, a logo, a fabric texture can shift between shots even when the face and body match, especially across a full angle or lighting change. Body proportion sits in between: generally stable at a similar distance from camera, but prone to drift on extreme perspective shots like a low-angle hero shot or an aerial. Knowing which element is weakest for a given shot type is what tells a supervisor where to look first during review, instead of eyeballing the whole frame every time.

How does a production team grade a shot before it ships?

Every candidate generation gets checked against three questions: does the face still match the locked reference, does the motion read as physically intentional rather than glitchy, and does the costume hold its texture and silhouette. A take that fails any one of the three goes back into the regeneration queue rather than shipping with a visible flaw patched in post. That three-point check, repeated 3-8 times per shot across 20-50 shots, is the actual work behind a video that looks like one continuous world instead of a slideshow of near-matches.