Skip to main content
AI characters drift between shots because each generation starts from scratch: if the character only exists in the prompt text, the model re-invents them every time. The fix is to define the character once as a reusable asset with a generated reference image, then pull that same asset into every shot instead of re-describing them. In Automat Studio that is what Character Studio and @-mentions are for.

​
Why drift happens

A text prompt is a fresh instruction. “A woman in her thirties with dark hair” describes a space of thousands of different people, and a video model picks somewhere in that space each time you ask. Add a new adjective in shot four and you have moved to a different part of the space — so shot four shows a different person. Nothing in a single prompt carries an identity forward. Consistency has to come from reusing a fixed reference, not from writing a better description.

​
Step 1: Define the character once, in detail

Create the character before you generate any shots. Give them a name, an age, and a description that pins down what the model would otherwise invent:
  • Face and build — hair colour, length and style; eye colour; height; build; skin tone; age markers.
  • Distinctive features — scars, glasses, freckles, tattoos, the things that make a face recognisable.
  • Default wardrobe — what they wear when nothing else is specified.
  • Bearing — how they hold themselves, which affects how they read on screen.
Vague descriptions are the single biggest cause of drift. “Tall with an angular face, cropped bleached hair and a chipped front tooth” survives regeneration; “attractive young man” does not. You also select the character’s voice here, which is what gets used later when you generate their dialogue — so the voice stays consistent too.

​
Step 2: Lock a reference image you are happy with

Generate the character’s image and keep regenerating until one is right, then make it the active generation. That image is the visual anchor for the character across the film, so it is worth spending credits to get it right before you shoot anything. Use the Draft quality setting while you are exploring and Final once you are close.

​
Step 3: Mention the character in every shot

In Shot Studio, type @ and pick the character to pull them into the shot rather than describing them again. The same works for props, locations and wardrobe — see mentioning assets. This is the habit that matters. Any shot where you type the description by hand instead of mentioning the asset is a shot that can drift.

​
Step 4: Let the aesthetic carry the look

Character consistency is only half of visual continuity — the other half is grade, lighting and texture. An aesthetic applied at project level keeps those consistent across every generation, so shots cut together as though one crew shot them. Six aesthetics are built in, and Independent and Hollywood plans can build custom ones from reference images.

​
Step 5: Anchor motion to a still

For a shot where the character’s face really matters, generate a first frame you are happy with and then turn that frame into video. Image-to-video starts from your approved still, so the face in the moving shot is the face you signed off — which is more reliable than generating video directly from text. If you need a specific performance, you can go further and upload your own filmed take with performance to video: the motion comes from your real performance and the AI restyles it.

​
Wardrobe and props drift too

The same rules apply. Define wardrobe and props as assets, link wardrobe to a character where relevant, and mention them per shot. A consistent face in an inconsistent jacket still breaks the cut.

​
What to expect

This approach gets you a recognisably consistent character across a film, which is what narrative work needs. It is not frame-perfect identity lock: current models still vary, and shot attributes such as lens and shot type steer a generation rather than dictate it. Budget for iteration on the shots where the face is large in frame, and use the reference-anchored image-to-video path for those.

​
Next steps