Ask any generator for the same person twice and you get siblings. Consistency is the hardest problem in generated film, and it is not solved by better prompts. Models have no memory between calls; every request starts from nothing. So consistency has to be engineered into the inputs, and every technique that works is a variation of one idea: decide what the character looks like once, in a form the model can copy, and give it that form every time.

Why it drifts

A text description is a wide net. "Woman, thirties, dark hair" matches millions of faces, and the model picks a different one each time. Even a long description leaves room: the shape of the jaw, the cut of the jacket, the exact shade. Lighting, angle and style pull the result further. Video models make it worse, because they also have to keep the face stable across frames and they trade fidelity to the reference for motion quality. Without engineering, the viewer needs subtitles to know who is who.

Lever one: references

Give the generator images, not adjectives. The character bible stores up to four reference faces and passes them to every scene that lists the character. In Story to film, the director draws a cast sheet per character before any scene exists: full body, neutral pose, plain background, in the film's style. A sheet works because an image pins down everything a description leaves open. This alone removes most of the drift.

A written look travels with the image: species or age, face, hair, costume, colours, one distinguishing mark. It is repeated in every scene prompt, so even a scene that fails to read the image still describes the right character. Write it for a model: concrete nouns and colours, not "beautiful".

Lever two: variants

A single front-facing reference does not help a profile shot. Pose variants, generated once from the bible, cover the angles your scenes need: full body, waist up, alternate outfits. Upgrading the bible adds variants without rebuilding it. The bible is the most expensive single operation in the studio, so build it once for characters who appear in more than two scenes and reuse it across films; for a walk-on, one reference image is enough.

Lever three: anchor frames

Approve a storyboard frame where the character looks right, then generate the video from that frame. The clip inherits the face instead of reinventing it. This is the keyframe-first path, and we use it at every quality tier: the image model composes the still from the sheets and the prompt, and the video model animates the still. Image models follow references far better than video models, and a failed frame costs a fraction of a failed clip.

Continuity frames are the same lever applied between scenes. When a scene continues the previous one in the same place, the clip starts from the last frame of the previous clip, and the character, the light and the set carry over. Use it only for true continuations; a new location needs a new frame or the old set leaks in.

What did not work

  • Provider reference endpoints for video. Acceptable for human faces, weak for animals and objects, and liable to be removed without notice. We routed around them; the keyframe path only needs image-to-video, which every provider offers.
  • Names in prompts. The model has never seen "Leyla" and treats the name as noise. The look has to be spelled out.
  • Very long prompts. Past a few hundred words, models weight the beginning and ignore the rest. Put the look first.
  • Merged cast entries. A sheet showing "the animals" cannot anchor a frog and an owl separately. One being per cast member, enforced.

How to tell whether it worked

Pick three scenes with the same character and compare face, costume and palette side by side. Score each 0 to 2: identical at a glance, recognisable with effort, could be someone else. Below four out of six, regenerate the sheet before regenerating any scene; the fault is almost always in the anchor. A cheaper proxy is the subtitle test: watch with the sound off and ask whether you can tell who is speaking. If you need the name on screen, consistency failed.

Fixing a character after the fact

Open Characters in the sidebar; the film's cast is saved there with its sheet. Add another reference image, a regenerated sheet or a different angle, or replace the master. Back in the film, regenerate only the scenes that use the character. Characters are yours after the film: attach them to scenes in any other film, including screenplays you write yourself.

Frequently asked

Does a photo of a real person work as a reference? Only for photorealistic styles, and only with the person's consent. For drawn styles, generate the sheet in that style; a photo drags realism into every scene.

How many reference images do I need? One good full-body sheet is enough for most films. A second image with a close-up of the face helps for scenes that are mostly faces, and the bible holds up to four.

Can two characters share a scene reliably? Yes, if both sheets reach the frame step and the prompt names them by position, such as the frog on the left and the owl on the right. Three or more in one frame is where current models start to blend features.