The drift problem
Generate a character from a prompt twice and you get two different people who roughly match a description. Generate twenty shots that way and your lead changes hair, age and jawline every cut. That is fine for a single image and fatal for a film, where the audience needs to recognise the same face from scene to scene.
Identity lock, step by step
Previzly treats a character as an identity, not as a prompt. You create the character once, with reference images and structured attributes: hair, eyes, build, height, skin tone, wardrobe and a colour palette. From that, the app generates a canonical avatar and stores the seed it was rendered with. Seed and avatar become the locked identity, and every keyframe that includes the character is rendered from them. The seed is visible on the character as a short code, so you can tell at a glance that two shots share an identity.
One voice per character
The same idea applies to sound. Upload a voice sample of about ten seconds and the character gets a cloned voice. Every dialogue beat spoken by that character is generated in that voice, with a delivery tone you set per line, from a whisper to furious. Lip-synced talking video is then built from the shot’s keyframe and the generated lines, so the face and the voice stay attached to the same identity.
Keeping it consistent while you edit
Consistency is not a one-time render. Previzly tracks the relationship between the script and the media. Change a shot’s action, camera or a line of dialogue and the shot is marked stale, which tells you precisely which frames no longer match without forcing you to regenerate everything. Keyframe takes work the same way: generate several candidates for a shot, keep the one that reads best, and animate from it.
Honest expectations
Identity lock makes consistency the default, not a guarantee. Video and image models are probabilistic, so a take can still miss, which is why takes exist and why the price of a keyframe is kept low. Characters need an avatar before their shots can render keyframes, and a cloned voice before their lines can be voiced; the app tells you when a shot is waiting on either. Reference images should show the face clearly and from more than one angle.
Where to start
Make the cast first, then the breakdown, then the keyframes. Casting first is the single habit that produces consistent films, because every shot is written for identities that already exist. If you want to see the whole workflow, read the guide on AI storyboard generation, and check the pricing page for what avatars and voice clones cost.