Anyone who has tried to make a short drama with AI video tools runs into the same wall: shot one looks great, shot five looks like a different cast. The face drifts, the jacket changes colour, the lighting decides it belongs to another film. Fixing it in post is painful, and re-rolling until it matches is expensive.
I've been shooting a 12-episode vertical drama with AI generation for the last few months, and the workflow below is what finally made characters hold together across 40+ shots. None of it is magic — it's mostly refusing to let the model improvise.
1. Write the character as a spec, not an adjective
"A tired detective in his 40s" is a prompt, not a character. The model will interpret it differently every time. What works is a list of fixed attributes you paste into every shot:
- age range and build
- hair length, colour, texture
- one signature garment (colour + material)
- one prop that never leaves the frame
- default lighting and lens
Keep it in a text file. It should be boring and identical every single time.
2. Lock the reference frame before you animate
Generate stills until one frame is exactly right, then treat that frame as canon. Every later shot starts from an image-to-video pass using that frame or a crop of it. Text-to-video is for exploration; image-to-video is for continuity.
If a tool can't start from an image, it can't carry a series. This is the single biggest filter when picking a generator.
3. Keep the camera language consistent too
Viewers forgive a lot, but they notice when the lens changes personality. Pick two setups — say a 35mm medium and an 85mm close-up — and stay inside them for the whole episode. Consistency in framing disguises small inconsistencies in the face.
4. Generate motion, not performance
Long prompts with emotional direction tend to produce drifting faces. Short prompts about motion ("she turns her head, wind moving her hair, slight handheld sway") keep the identity stable and give you room to act in the edit.
5. Cut on motion, not on stillness
AI clips are strongest in the middle. Trim the first and last half second, where morphing usually shows up, and cut while something is moving. Two 3-second clips cut on a gesture read better than one 6-second clip that slowly melts.
What I actually use
For the drama I've been running the image-to-video path in Pavo AI, mainly because the character stays recognisable between shots and the vertical output is already framed for phones. The prompt library at pavoai.app/prompts is a decent starting point if you're building your own shot list — I copied the structure, then replaced the descriptions with my own character spec.
The rest of the stack is unglamorous: a spreadsheet for shot IDs, a folder per character, and a rule that nothing gets generated until the reference frame is approved.
The takeaway
Consistency isn't a model feature you buy once. It's a production habit: fixed character specs, locked reference frames, a narrow set of camera setups, and short motion prompts. Do that and even a modest generator will hold a series together. Skip it and the best model on the market will still give you twins.
Top comments (0)