From face to film: AI video in three stages

Most people use AI video exactly the wrong way round. They open a text-to-video tool, describe a scene, hit generate and keep trying until something usable comes out. Then they do the same for the next shot — and suddenly the character has a different nose.

The fix is not a better prompt. It is a different order. You build the person first, then the images, then the motion. Three phases, where each phase locks in decisions so the next one has less room to go wrong.

Phase 1 — Build the character

Everything that follows depends on this. If the identity is not locked in, you spend the rest of the project fighting for likeness instead of directing.

Gather the reference material

You need twenty or more photos of the same person. Their quality determines the quality of everything after, so be strict:

• Well lit, no harsh shadows. Hard side light teaches the model half a head.
• No sunglasses, no caps, no cropped faces. Anything that hides the features gets baked into the identity as vagueness.
• Vary angle and expression. Front-on, three-quarter, profile. Neutral, smiling, mid-sentence. A model trained on twenty near-identical selfies learns one pose, not a face.
• Include at least one full-length photo. This is the step everyone skips, and the reason AI characters often have slightly wrong body proportions.
• Use recent photos — ideally from the last four or five months. Older references give you a likeness of who someone used to be.

Training

Upload the set and train. That takes roughly three to five minutes. What you get back is a persistent identity: a trained layer that carries the person's features into every generation, regardless of lighting, camera angle or style.

Check before you build on it

Do not go straight into production. Generate a grid of six to eight images in which you only change the camera angle — same clothes, same light, same setting. Put them side by side. If the face holds up, your identity is solid. If it drifts on the profile shots, retrain with more material shot from an angle.

Ten minutes of checking saves you from discovering the problem only after you have made forty shots.

Write the character block

While it trains, write a short block that you paste into every prompt from now on:

CHARACTER: [name], [age range], [build/height]
CLOTHING: [specific garments, colours, materials]
FEATURE: [one distinctive detail — glasses, watch, scar, jacket]
DEMEANOUR: [how they carry themselves — quiet, restless, closed, open]

The trained identity takes care of the face. This block takes care of everything the face does not — and stops your character quietly changing jackets between shot two and shot three.

Phase 2 — Turn the story into stills

This is where the actual filmmaking happens, and it is the phase people rush. Every minute you spend here is worth ten in the animation phase: a bad image animates into a bad clip, while a good image animates into a good clip almost by itself.

Cut the story into beats, not scenes

A scene is ‘she arrives at the office’. A beat is one unit of change. Write your story as a list of beats, one line each:

• She stands outside, hesitates.
• She decides.
• She walks in.
• Everyone looks up.
• She does not flinch.

Five beats, five shots. If you cannot name what changes in a beat, it does not need a shot.

Give every beat a shot size

Shot size carries meaning, so do not choose at random. The reliable pattern:

Type of beat	Shot	Why
Establishing	Wide	Tells you where you are
Anticipation	Detail / macro	Holds information back, builds tension
Action or reveal	Medium	Body language reads clearly
Emotion or release	Close-up	The eyes carry the moment
Consequence	Wide again	Shows what has changed

Alternate them. Two wides in a row is dead air; two close-ups in a row is claustrophobic. The rhythm of changing shot sizes is your pacing.

Write the image prompts

Every prompt has four parts, in this order:

• Frame and composition — medium shot, subject on the left of frame, shallow depth of field
• Character block — pasted from phase 1
• Setting — location, time of day, three concrete props
• Light — source, direction, hardness, colour temperature

Light is the parameter that separates amateur from cinematic, and it is exactly the one almost everyone leaves out. ‘Hard low sunlight through a venetian blind, warm, strong shadow pattern on the back wall’ gives you a different film from ‘office, daytime’.

Select ruthlessly

Generate four to six variants per beat. Keep one. Look out for:

• Face — no melted ear, no wrong eyeline
• Hands — still the quickest giveaway
• Continuity — same jacket, same watch, same haircut as the neighbouring shots
• Room to move — leave space where the movement is going

That last one matters more than it sounds. If you are planning a push-in, frame slightly wider than you want the final image, so the move has somewhere to land. If the subject walks to the left, leave room on the left-hand side.

Put all your keepers in order before you move on. This is your storyboard, and the last cheap moment to notice that beats three and four say the same thing.

Phase 3 — Turn the images into motion

Only now do you animate, image-to-video, with one approved image per clip. Because the frame is already locked, the model has just one job left: moving the camera. That is a much smaller task than inventing a scene, and that is exactly why the result is more stable.

One camera move per shot, chosen deliberately

Use named camera presets instead of describing the move in prose. Presets are modelled on real camerawork and so carry believable mass, instead of the floaty drifting that gives AI video away.

Match the move to the function of the beat:

• Slow dolly in — we are getting closer to the truth. For establishing shots and emotional build-up.
• Macro push — anticipation. Strong on detail shots.
• Crash zoom or whip pan — the reveal. Use it once. Twice is a music video.
• Slow orbit — the release. Circling someone says: look at this person.
• Static with movement in frame — underrated. Locked-off camera, moving subject. Use it to make the moving shots hit harder.

One move per clip. Two moves fight each other, and the model resolves that conflict by doing both of them halfway.

Generate long, edit short

Render every clip at the full available length — usually five seconds — and use two or three of them in the edit. AI clips fall apart at the end far more often than at the start, so you want your surplus at the tail, where you can throw it away.

Cut on motion

This is the technique that makes a sequence read as filmed rather than as stapled-together generations. End clip A while the camera is still moving, and start clip B mid-move in a compatible direction. The eye reads continuous motion across a cut as continuous coverage of a single event.

Cutting from a still frame to a still frame does the opposite: it exposes every seam.

Sound, then lip sync, then stop

Add ambient sound and music before you start on dialogue. Most sequences that ‘feel off’ suffer from silence, not from bad images.

If the character has to speak, sync one shot. One line that lands on your closing close-up does more than twenty seconds of continuous AI dialogue, which reliably tips over into the uncanny. Restraint is the whole trick.

Budget discipline

Heavy models burn through credits fast, and the temptation is to run everything on the best model straight away. Don't. Work on a cheap model until frame and camera move are exactly right, and only then render your final choices at full quality. You typically make three to five times as much material as you use — budget for that ratio, not for the final running time.

In short

• Train the identity properly, check it, and write a character block that you paste into every prompt.
• Cut the story into beats, alternate shot sizes, and select images until every frame is one you would want to pause on.
• Animate with one move per shot, generate long, cut on motion, and use dialogue once.

The counter-intuitive part is that almost none of this is about video. It is about doing the photography and the storyboard first — the same order a real production works in. The tools got faster. The order did not.