How to restyle real footage with AI and keep it watchable

Video-to-video restyling is the one AI task that takes something you already shot and gives it back changed. The motion stays, the framing stays, the timing stays; what changes is what the frame looks like. Snow where there was sun, night where there was afternoon, watercolour where there was a phone.

The demos are extraordinary and honest. The failures are equally real and rarely shown. This is a guide to the space between them: which changes hold up, which quietly destroy the footage, and how to run a scene through it without the audience noticing the seams.

What restyling is good at

The rule is that restyling is strongest when it changes the world and weakest when it changes the subject.

Changes to the world hold up beautifully:

• Weather and season. Rain, snow, fog, autumn leaves. The model adds it consistently across the motion, and because the ground truth (the street, the building, the person walking) is still there under it, the result is stable.
• Time of day. Afternoon to dusk, day to night. Lighting changes are what these models learned first and do best.
• Materials and surfaces. A brick wall to marble, a wooden table to glass. Fine as long as the geometry stays.
• Global style. Painterly, illustrated, film stock, anime. When the whole frame goes through the same transformation the result is coherent, and the more stylised the target, the more forgiving it is of the small errors underneath.

Changes to the subject are where it breaks:

• Faces. A restyle that touches a face redraws it slightly differently on every shot. The person is recognisably the same in shot one and recognisably not in shot three. Nothing about the shots being "the same person" is known to the model.
• Text and logos. Signs, labels, screens. Redrawn, and redrawn wrong. If a shot contains a word the audience should read, either protect it or expect to replace it.
• Hands and fine mechanics. Fingers, cutlery, buttons. Same story as generation: the most-often-wrong parts of a frame are the most-often-wrong parts of a restyle.
• The subject's identity. "Make the person a knight" works in a still and drifts in motion. The armour will be different armour in every shot.

So the question before restyling any scene is: is the thing I want to change the world or the subject? World changes are a restyle. Subject changes are usually a generate-from-scratch job, or a grade, or a cut.

Consistency is the whole game

A single restyled clip is easy. A scene is hard, because a scene is several shots that must look like the same place, and a restyle is run per shot with no memory of the others.

The way to get consistency is to remove every variable except the footage:

Same model, same prompt, same settings, same sitting. Write the prompt once, run every shot in the scene with it verbatim, on the same model at the same strength, in one session. Do not "improve" the prompt on shot three. If shot three needs a different prompt, all the shots need it; go back and re-run the lot.

Describe the world, not the shot. A prompt that says "heavy snow, overcast, late afternoon, blue-grey light" transfers across a wide, a medium and a close. A prompt that says "a woman walks past a snowy shop window" is specific to one shot and will pull the others toward it.

Grade them together afterwards. Even a consistent restyle leaves shots that are a touch warmer or darker than their neighbours. A colour grade applied to the whole scene, with a small per-shot correction, pulls them into one look; it is the same step a real shoot needs, for the same reason.

Cut hard, dissolve never. A dissolve between two restyled shots shows both versions of the world at once, which is exactly where their differences show. Cuts hide what dissolves reveal.

Split first, and keep it short

A restyle bills on the length of clip you send, and several models will not accept more than about ten seconds; a couple also need footage at least 720 pixels tall, and both refusals happen before anything is charged. So split the clip to the section you will use before running it, and expect to send shots, not scenes.

Short also helps the result. The longer the clip, the more the style drifts across it; a six-second shot holds a look that a twenty-second one loses.

Draft cheap, finish good

The models on this task are priced very differently, some flat per run and some per second, and the number under the button is worth a glance before every run. The economical sequence is to try the prompt on a cheap model, on one shot, until the idea is right; then run the whole scene on the model whose output you want to ship. What you are paying for on the expensive model is polish, and there is no point paying for polish on an idea you are still changing.

The result always lands as a new clip in the library; the original is never overwritten, and replacing it on the timeline is one undo away.

When a grade is the honest answer

A lot of the time, what someone wants when they say "restyle" is a look: warmer, colder, filmic, faded, teal-and-orange, high contrast. That is not a restyle. It is a grade, and a grade is better on every axis: instant, free, exactly consistent across shots, and incapable of changing a face.

The test is whether the change moves pixels around or only recolours them. Snow moves pixels. "Make it moody" recolours them. If the answer is recolour, grade it and keep the restyle for when the world genuinely has to change.

Between the two sits the Looks task, a fixed transformation applied without a prompt; it is what to try when a grade is not quite enough and a free-prompt restyle is more than you need.

Stills first when you can

If a scene is going to be restyled heavily, a shortcut that saves both money and disappointment: restyle one frame of it as an image edit first. You will know in one cheap run whether the look you want is achievable and whether the subject survives it. If it does, restyle the clips. If it does not, you have learned it at the price of an image rather than a scene.

For a look you will use again and again, across projects, a trained style on images is the more reliable route than re-prompting a video restyle each time.

The order that works

1. Ask whether the change is to the world or the subject. World: restyle. Subject: something else.
2. Test the look on a single frame as an image edit before touching video.
3. Split every shot to what you will use; keep them short.
4. Draft the prompt cheap on one shot, then run the whole scene with one prompt, one model, one sitting.
5. Grade the scene together afterwards to close the last gaps.
6. Cut, never dissolve, between restyled shots.