FLUX Video Edit [fast]

Introduction

Removal is the edit that gets asked for most, and it is rarely about art direction. A property clip has the previous owner's cleaning products on the counter. A storefront b-roll has a bicycle nobody cleared. An evergreen packshot has a promotional sticker that expired in March. The footage is fine and one thing in it cannot ship.

Removal in FLUX Video Edit [fast] is a prompt like any other edit, with one extra demand: the model has to invent what was behind the thing you took out, and do it consistently for every frame.

The camera keeps moving, the daylight keeps falling the same way, and the three things that had to go are gone. The lemons, the eucalyptus and the cookbooks were never mentioned, so they stayed. This guide covers how to phrase a removal, why a moving object is harder than a parked one, and the cases where removal is the wrong request.

The request

Nothing about a removal is structurally different. The clip goes in inputs.video and the instruction goes in positivePrompt.

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: 'Remove the black bicycle parked against the wall on the right, and continue the bakery wall and the pavement behind it.',
  inputs: {
    video: 'https://vm.runware.ai/video/os/a14d18/ws/2/vi/8f1d3c67-2b94-4e50-a3f7-6c9d2b5e1a84.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": "Remove the black bicycle parked against the wall on the right, and continue the bakery wall and the pavement behind it.",
            "inputs": {
                "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/8f1d3c67-2b94-4e50-a3f7-6c9d2b5e1a84.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": "e93b7d21-4c58-4a06-b71f-8d2c5a9e3f10",
      "model": "bfl:flux@video-edit-fast",
      "positivePrompt": "Remove the black bicycle parked against the wall on the right, and continue the bakery wall and the pavement behind it.",
      "inputs": {
        "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/8f1d3c67-2b94-4e50-a3f7-6c9d2b5e1a84.mp4"
      },
      "deliveryMethod": "async"
    }
  ]'

runware run bfl:flux@video-edit-fast \
  positivePrompt="Remove the black bicycle parked against the wall on the right, and continue the bakery wall and the pavement behind it." \
  inputs.video=https://vm.runware.ai/video/os/a14d18/ws/2/vi/8f1d3c67-2b94-4e50-a3f7-6c9d2b5e1a84.mp4 \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "e93b7d21-4c58-4a06-b71f-8d2c5a9e3f10",
  "model": "bfl:flux@video-edit-fast",
  "positivePrompt": "Remove the black bicycle parked against the wall on the right, and continue the bakery wall and the pavement behind it.",
  "inputs": {
    "video": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/8f1d3c67-2b94-4e50-a3f7-6c9d2b5e1a84.mp4"
  },
  "deliveryMethod": "async"
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "e93b7d21-4c58-4a06-b71f-8d2c5a9e3f10",
    "videoUUID": "4b7e2a95-6d13-4c80-9f52-1a8c3d7e5b26",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/4b7e2a95-6d13-4c80-9f52-1a8c3d7e5b26.mp4"
  }
]

There is no mask, and there is no way to supply one. The noun phrase is the mask, which is why the rest of this guide is about writing it.

Name the target and what continues behind it

A removal instruction has two halves. The first names what goes. The second names what the model should put in its place, and it is the half people leave out.

"Tidy the kitchen" leaves both halves to the model. Tidying is a judgment, and its idea of what counts as clutter includes things the stylist put there deliberately:

The anchored version removes three things because it named those three things, and the lemons, the eucalyptus and the cookbooks survive because it did not. The counter reads as continuous because the prompt said what the counter is. "Continue the pale quartz counter surface" is a small phrase doing a large job: without it the model still has to fill the hole, and it fills it with whatever it infers, which on a busy frame can be a fourth object.

Naming the fill matters most when the surface behind the target is patterned, textured or lit unevenly. A plain white wall forgives a vague instruction. A wood grain, a tiled backsplash, a rug pattern or a gradient of daylight does not, because the eye tracks continuity across frames and a fill that drifts is obvious in motion even when a still frame looks fine.

Static objects and moving objects

Two removals from the same clip are not equally hard. This storefront has a parked bicycle on the right and a courier walking across the middle:

The bicycle covers one fixed patch of wall. With a locked-off camera, the model has to invent that patch once and hold it, and nothing in the frame ever contradicts the invention.

The courier covers a different patch in every frame. The fill has to be invented continuously, and each new frame has to agree with what the previous one established about a piece of window the camera has never actually seen. When a moving-object removal fails, the failure is usually a shimmer in the fill rather than a ghost of the subject, so watch the background he crossed rather than the space he vacated.

Two things make a moving removal easier, and both live in the source rather than the prompt:

• A locked-off camera. A pan or a dolly means the background behind the subject is being invented and reprojected at the same time.
• A background the clip reveals elsewhere. If the courier walks past a stretch of window that is unobstructed earlier in the shot, the model has seen the answer. Choosing the take where the subject crosses an area the camera also shows empty is worth more than any phrasing.

Removing people

A person is a moving-object removal with two extra complications: contact shadows and reflections. The body is easy to name and hard to erase completely, because it leaves evidence on surfaces around it.

Name the evidence when it is visible:

For a person on a reflective floor or in front of glass, extend the instruction to cover what they leave behind: "and remove their shadow on the pavement", "and remove their reflection in the window". A removal that clears the body and keeps a shadow reads worse than no removal at all, because a floating shadow is a mistake nobody makes accidentally in real footage.

Removing stickers, badges and graphics

The commercially useful removals are often the small flat ones. Promotional roundels, sale badges, price stickers and third-party logos all date a piece of footage, and a packshot that is otherwise evergreen has to be reshot every season because of them.

Two details make this case behave better than a general removal. The sticker sits on a surface whose design the model can see the rest of, so "continue the label artwork underneath it" points at evidence that is actually in frame. And the bottle rotates, which means the sticker's patch is revealed by the motion itself on the frames before and after.

The same instruction shape covers a logo on a laptop lid, a badge on a garment, a brand mark on a coffee cup and a license plate on a car. Say what the surface is, not just that the graphic should go.

When the background is never revealed

Every removal is a bet that the model can reconstruct what was hidden. That bet gets thin when the object is large, static and never moves off what it covers. This box sits against a panelled wall for the entire clip on a locked-off camera, so nothing in the footage ever shows what the panelling does behind it:

The reconstruction has to invent panel spacing, a skirting line and a floor reflection that nothing in the source constrains, and the panel rhythm it invents has no reason to match the rhythm visible either side of the box.

Replacing is the cheaper bet. A basket occupies roughly the same footprint, so the model reconstructs a much smaller area and the parts it invents sit behind an object the eye accepts. When a removal has to reconstruct a large patch of structured background, ask what could plausibly stand there instead, and put that in the frame.

The general rule is worth holding onto: the difficulty of a removal is the area and structure of what it uncovers, not the size of the thing removed. A large object against a plain wall is easy. A small object sitting on a patterned rug in front of a bookshelf is not.

Tips

1. Write both halves. Name the target, then name the surface that continues behind it. The second half is what keeps the fill stable across frames.

2. Never ask for a judgment. "Tidy", "clean up" and "declutter" hand the model a decision about what belongs, and it will remove things a stylist chose on purpose.

3. Point at the object by appearance and position. "The black bicycle against the wall on the right" resolves. "The bike" resolves only when there is exactly one.

4. Say "for the whole clip" on anything that moves. It is the difference between a subject who was never there and one who fades out halfway.

5. Prefer locked-off takes for moving-subject removals. A static camera means the background is invented once instead of invented and reprojected.

6. Pick the take where the background is revealed elsewhere. If the shot shows the wall empty at some point, the model has seen the answer it needs.

7. Name the shadow and the reflection. A cleared body with its shadow still on the floor is a worse frame than the original.

8. Say what the surface is when removing a graphic. "Continue the label artwork underneath" beats "remove the sticker" on anything with a design behind it.

9. Replace instead of removing when the hidden area is large and structured. An object the eye accepts is easier to render than a reconstruction of panelling nobody has seen.

10. Judge the result on the background, not the gap. Removals fail where the fill meets the real footage, so watch the edges and the frames either side of the motion.