FLUX Video Edit [fast]

Introduction

Animation is the most expensive way to say something simple. A thirty-second explainer in flat vector style is weeks of storyboarding and keyframing, and the thing it is animating is usually a person doing something a camera could have filmed in an afternoon.

A restyle inverts that. You film the afternoon, then change the medium and keep the performance. The timing, the camera, the gesture and the beat where he steps back are all things a human already got right, and none of them have to be animated by hand.

He draws on the same beat and steps back at the same moment. This guide covers how to name a style the model can actually render, how far a restyle reaches into the frame, and when a restyle is the wrong tool.

The request

Restyling is the same call as any other edit. The instruction is what makes it whole-frame rather than local.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'bfl:flux@video-edit-fast',
  positivePrompt: 'Restyle the whole shot as cel-shaded anime, with hard-edged shadow blocks, clean ink linework and desaturated flat backgrounds.',
  inputs: {
    video: 'https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a8f2c61-7d54-4b90-a2e6-8c1d5f3a9b47.mp4'
  },
  deliveryMethod: 'async'
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "bfl:flux@video-edit-fast",
            "positivePrompt": "Restyle the whole shot as cel-shaded anime, with hard-edged shadow blocks, clean ink linework and desaturated flat backgrounds.",
            "inputs": {
                "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a8f2c61-7d54-4b90-a2e6-8c1d5f3a9b47.mp4"
            },
            "deliveryMethod": "async"
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "b41c8e57-2a96-4d03-8f71-5e2a9c4d7b18",
      "model": "bfl:flux@video-edit-fast",
      "positivePrompt": "Restyle the whole shot as cel-shaded anime, with hard-edged shadow blocks, clean ink linework and desaturated flat backgrounds.",
      "inputs": {
        "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a8f2c61-7d54-4b90-a2e6-8c1d5f3a9b47.mp4"
      },
      "deliveryMethod": "async"
    }
  ]'

runware run bfl:flux@video-edit-fast \
  positivePrompt="Restyle the whole shot as cel-shaded anime, with hard-edged shadow blocks, clean ink linework and desaturated flat backgrounds." \
  inputs.video=https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a8f2c61-7d54-4b90-a2e6-8c1d5f3a9b47.mp4 \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "b41c8e57-2a96-4d03-8f71-5e2a9c4d7b18",
  "model": "bfl:flux@video-edit-fast",
  "positivePrompt": "Restyle the whole shot as cel-shaded anime, with hard-edged shadow blocks, clean ink linework and desaturated flat backgrounds.",
  "inputs": {
    "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a8f2c61-7d54-4b90-a2e6-8c1d5f3a9b47.mp4"
  },
  "deliveryMethod": "async"
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "b41c8e57-2a96-4d03-8f71-5e2a9c4d7b18",
    "videoUUID": "e27d4a93-5c18-4f60-b834-1a7c9e2d6b50",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/e27d4a93-5c18-4f60-b834-1a7c9e2d6b50.mp4"
  }
]

Name a style, not a mood

This is the one place where more words reliably help, because the model is being asked to re-render every pixel and needs to know what the new pixels are made of. "Make it artistic" is an instruction with no content:

The instruction that works carries three layers, and it is worth writing all three every time:

• The medium. Flat 2D vector animation, cel-shaded anime, hand-painted watercolor, technical line drawing, stop-motion clay, 16mm film. A named medium brings a whole set of conventions with it.
• The marks. How the medium makes an edge and a shadow: thick even outlines, hard-edged shadow blocks, loose ink contours, visible paper texture, no gradients. This is what separates two styles that share a medium.
• The palette. Either specific colors or a described range. Left unsaid, the model keeps something close to the source's colors, which often fights the style you asked for. Flat vector work with photographic color looks like a filter.

The strongest phrasing goes one step further and names what replaces what. Black Forest Labs write a live-action conversion of an animated clip as "Fine velour fibers and sewn seams replace the ink outlines", which tells the model both the new surface and the old mark it is displacing. A restyle written as a set of substitutions is more precise than the same restyle written as a description, because it points at what is actually in the source frame.

One clip, several looks

Because the source carries all the motion, a single recording can be delivered in as many styles as a campaign needs, with identical timing across every version:

This is the pattern behind a brand that ships one message in several visual identities, or a marketplace that offers the same stock footage in multiple treatments. The cut is identical across versions, so they drop into the same edit and the same sound mix without retiming anything.

What survives a restyle

A restyle is the most aggressive edit this model does, and the split between what it keeps and what it discards is sharp.

Kept: the camera path, the framing, the timing, the blocking, the gestures, the duration and the audio. Every temporal property of the shot is untouched, which is exactly why the technique is worth using.

Discarded: photographic texture, real material response, and most of what makes a specific face that face. A restyle re-renders the frame in a medium that does not have skin pores or specular highlights, so it cannot preserve them.

Styles closer to photography keep more. A film-stock treatment is a restyle in the same sense, but it changes how the image was captured rather than what it is made of, so faces, materials and detail all survive it:

The useful mental model is a spectrum of distance from the source. Film stock and photographic-era looks sit close to it and keep identity. Painterly and illustrated styles sit far from it and trade identity for medium. Pick knowingly, because the ones in the middle are where a result most often disappoints: a style strong enough to lose the face but not strong enough to look deliberate.

Restyling part of the frame

A photoreal presenter in an illustrated world is the thing everyone wants, and the word restyle is what stops you getting it. Restyle is an instruction about how the frame is rendered, so it applies to the frame. Ask for the same result as a background replacement and the boundary holds:

Two attempts phrased as restyles converted the man along with the room, one of them after explicitly saying not to and spelling out what photoreal meant. The version above never uses the word. It replaces the background, which is an object-level edit, and gives the two halves equal detail: the palette and marks for the new background, and the skin, hair and photographic lighting for the man.

The rule generalizes past this one shot. Pick the verb that matches the scope you want. Restyle converts the picture. Replace the background converts a thing in it. A request for a styled world around an untouched subject is a background job wearing a restyle's clothes.

Restyling for a different purpose

Not every restyle is a stylistic choice. A style change can also be a functional conversion, and the technical registers are some of the most useful:

One rotation covers a whole documentation set: the same packshot becomes an assembly diagram, a patent-style illustration, a coloring-page asset or a schematic. For anything where the geometry matters and the finish does not, a line-drawing restyle carries more information than the photograph did.

Finishing a blockout

The substitution phrasing has a use that goes past styling: a gray previz render can be restyled into a finished shot. The blockout supplies the layout, the camera move and the timing, all of which are cheap to build in a 3D scene and expensive to describe to a text-to-video model. The restyle supplies everything else.

Write it as a mapping, one placeholder at a time: this gray box becomes that shelving unit, this slab becomes that counter, this featureless figure becomes that person. A blockout has no texture for the model to reason from, so an unmapped element is an element it will guess at, and the guess is usually the least interesting object that fits the silhouette.

Then say explicitly that the camera and the layout are unchanged. This is the one place where a preserve clause earns its keep, because the instruction is asking for so much that the model has every reason to treat the geometry as negotiable too.

The same technique covers a previz chase cut for a film, a walkthrough of an unbuilt space for a property listing, and a level graybox dressed into concept art. In each case the expensive thing is the motion through a space, and it was authored deliberately rather than prompted.

Restyle or generate

There is a real line here, and it comes down to what you are keeping.

Restyle when the performance is the point. A gesture that lands, a camera move that works, timing that was cut to music. Those are expensive to reproduce and free to keep.

Generate when only the idea is the point. If the shot is a generic subject doing a generic thing, a styled model asked for that directly gives you a cleaner result, because it is drawing the frame rather than translating one.

The hybrid is worth knowing about: film the motion you need as a rough reference, then restyle it. A phone recording of the right gesture, at the right speed, restyled into your target medium, is a faster route to a specific piece of animated motion than describing that motion to a text-to-video model and hoping.

Tips

1. Write medium, marks and palette. Three layers in one sentence is the difference between a style and a filter.

2. Never ask for a mood. "Artistic", "stylish" and "cinematic" describe a reaction, not a way of making an image.

3. Prefer flat styles for stability. Hard edges and solid fills hold across frames. Fine texture has thousands of marks with no reason to land twice.

4. Say the palette or accept the source's. Photographic color under a vector style is the most common reason a restyle looks half-done.

5. Expect identity to go. A strong restyle produces a character in the same position, not a stylised portrait of your presenter.

6. Use film-stock looks when faces must survive. They change the capture rather than the medium, so detail and likeness come through.

7. Never say restyle when you mean the background. Restyle converts the whole frame; "replace the background" keeps the subject photoreal.

8. Name what the subject touches when replacing a background. Props that belong to the room leave with it, and a presenter writing on empty air is the tell.

9. Restyle a rough reference for specific motion. Filming a gesture and converting it beats describing it to a text-to-video model.

10. Map a blockout element by element. Say what each gray placeholder becomes, then pin the camera and the layout. Anything unmapped gets guessed at.

11. Restyle every shot in the sequence. A converted clip cannot cut against untouched footage, and the mismatch is more obvious than with any other edit.