November 6, 2026 · 8 min read
Character Continuity in AI Video: Why Your Character Changes Between Shots
Your character's face changes because each shot is generated from near-zero memory. The fix is not a better prompt — it is an anchor, a descriptor block, and a card you never improvise on.
The short answer
Your AI character changes between shots because every shot is generated almost from scratch: the model holds no persistent memory of your character, so each new prompt re-rolls their face, wardrobe, and proportions from probability. The measured size of the problem: researchers scoring multi-shot character consistency found it collapses from 7.99 to 0.55 out of 10 when no visual anchor is used (Jain et al., 2025). The working fix is a discipline, not a prompt trick: a visual anchor (reference images the generator conditions on), a verbatim descriptor block pasted unchanged into every prompt, and a Character Lock Card that fixes both in one place for the whole series. ContentCraft's Continuity Kit is built on exactly that discipline — your first Lock Card is free, and each series carries a SHA-256-locked continuity manifest so the canon is verifiable, not remembered.
Why the drift happens
A video generator does not know your character; it knows the distribution of faces that fit your description. "A woman in her 30s with curly hair" is a crowd, and every shot samples a different member. Change the lighting words, the camera words, or the outfit words and the sample moves again. Three failure layers stack: identity drift (the face itself changes), attribute drift (wardrobe, hair, and props mutate because the prompt mentioned them differently), and interaction breakdown — two-character scenes remain, in the trade press's assessment, unsolved across all platforms as of mid-2026, because two drifting identities must also stay consistent relative to each other.
The anchor method, step by step
| Step | What you fix | Why it works |
|---|---|---|
| 1. Cast once | Generate or choose the reference images that ARE the character | An image anchor conditions generation far more strongly than adjectives |
| 2. Write the descriptor block | One verbatim paragraph: face, hair, build, signature wardrobe, palette | Identical words in every prompt remove attribute drift at the source |
| 3. Lock the card | Anchor + descriptor + voice/style notes on one Character Lock Card | The canon lives in one place the whole series — and any collaborator — can check |
| 4. Carry it over | The card's carry-over block pastes into every generation session | Continuity becomes a paste, not a memory test, across episodes and tools |
| 5. Verify the manifest | The series' SHA-256 continuity manifest records the locked canon | Disputes about "which version is canon" end with a hash, not an argument |
What this does not fix
Honesty about the ceiling: anchors and cards hold a character's identity; they do not solve two-character physical interaction, hands, or long continuous action — those remain generation problems no workflow layer removes. What the discipline removes is the unforced error: the solo-character series whose lead is a stranger every episode. For a series business, that is the difference between a character audiences can follow — literally, as a person they recognise — and a new stranger each week. A locked character still needs viewers on a schedule: the Traffic Machine plans one traffic activity per day around a series, so each episode lands in front of an audience instead of hoping. Build your first card in the Continuity Kit, run your scripts through the Viral Autopsy so the story holds as well as the face, and mind the December 2 disclosure deadline that applies to the published result.
Questions
Why does my AI character look different in every shot?
Because each shot is generated from near-zero memory of the last one: your text description names a distribution of possible faces, and every generation samples it again. Without a visual anchor, measured consistency collapses from 7.99 to 0.55 out of 10 (Jain et al., 2025).
What is a Character Lock Card?
The single canonical record for one character: reference anchor images, the verbatim descriptor block used in every prompt, and voice/style notes. In ContentCraft’s Continuity Kit the first Lock Card is free, and a series’ locked canon is recorded in a SHA-256 continuity manifest.
What is a descriptor block?
One fixed paragraph describing the character — face, hair, build, signature wardrobe, palette — pasted unchanged into every prompt for every shot. Identical wording removes the attribute drift that paraphrasing introduces.
Can continuity tools fix two characters interacting?
Not yet, honestly: two-character interaction is reported as unsolved across all AI video platforms as of mid-2026. Anchors hold each identity; the relative interaction between two generated characters remains a generation-level problem.
What is a continuity manifest?
A SHA-256-locked record of a series’ canonical cards and blocks. It makes the canon verifiable — anyone can check that the character definition in use matches the locked one — instead of relying on memory across episodes and collaborators.
Try it in the workshop.
Everything in this guide is a tool inside ContentCraft AI. Free to start.