Keep the original person or regenerate the scene?
To change a video background, first decide whether you need the original foreground footage or can accept a newly generated version of the scene.
A video background remover isolates the foreground for compositing over another image or clip. You keep using the original performance, but the mask can lose hair, fingers or parts of the subject. Inspect the edges throughout the shot. Adobe Express documents this background-removal route.
An AI video background changer using generative editing or reference generation can rebuild the location, lighting and performer. It can suit a creative remake, but requests to preserve the person and motion need checking. Runway describes a generative backdrop workflow; Alibaba Cloud distinguishes reference generation from video editing and character replacement. Those are method descriptions, not a ranking of tools we tested.
The example below uses FlyAIgh’s BAD template: a prepared video, a character reference and an optional scene reference. It shows how to move a character’s performance to a different setting, with changes to the character too.
Real example: office and rooftop variants
On the left is the office human sample used as the video reference. In the middle is a tabby version in the office. On the right is a tabby version on a sunset rooftop. The two cat clips were generated separately from the human sample; the rooftop run also used a new scene reference.
The rooftop variant replaces the visible office desks, screens and banner with a railing, skyline and sunset. The tabby’s white shirt and tie remain recognisable, and the broad arm movements and closing sequence are similar across the variants. The warm colour and floor reflections change with the setting.
The body, paws, clothing folds and foot positions still need individual review. Matching broad movements does not establish identical poses or timing. The references were prepared in separate runs, so this is not a controlled experiment where only the background input changed. It is also not an office cat clip with a new background pasted behind it.
What do the video, character and scene images provide?
- The video supplies the performance and framing to follow. BAD uses the office sample, even when its display preview shows a transformation from the mannequin.
- The character image describes the dancer’s appearance, body and outfit. Inspect the selected image after any reference extraction.
- The scene image supplies the new setting. Show a usable floor, a plausible horizon and the light you want; check that the camera angle suits the performance.
A scene image can contain a person. The BAD prompt asks the model to take only the setting from it, while the character image supplies identity. That separation is a request to verify in the result. If the references suggest different characters or scales, resolve the conflict before submitting.
When making a scene from text, Recast first edits a frame of the office sample with your selected character to prepare the new reference. Preview that still before asking for a video. After changing the character, prepare the scene reference again if the workspace marks it as outdated.
Change the BAD background in Recast
- Open BAD and play the sample. It is a fixed-camera, portrait 9:16 dance lasting 15 seconds; confirm this is the performance you want.
- Add the dancer reference and inspect it. BAD requires a character image; this workflow changes the dancer as well as offering a new scene.
- In the optional background step, upload a scene image or choose the text-description option. Our rooftop request described a city rooftop at sunset with distant buildings and evening light.
- If using text, generate the scene reference and check its floor, framing and character placement. Approve the visual choice before continuing to the video stage.
- Choose an available model and resolution, then read the prompt preview and video quote. Wan 3 at 720p produced this example; that is a case record, not a best-model claim.
- Generate, compare the full result with the sample, and download the version you want to edit. Check the scene and performer together before adding music.
Open BAD’s character and background workflow
Describe the new location and the old objects that should disappear. The recorded rooftop prompt asks to remove the office, monitors, banner and warning lights, and to follow the video’s movement and fixed camera. These instructions define the goal; they do not lock every frame. Make one change at a time when reviewing a revision.
Does the person belong in the new scene?
Watch at normal speed, then pause the source and result at the same moments. A convincing background in one frame does not show whether the feet stay grounded or the face survives a turn.
| Check | Compare | Review next |
|---|---|---|
| Identity and clothes | Face, fur, outfit and body proportions through turns | Character reference and conflicting figures in the scene image |
| Movement and framing | Gesture order, timing, scale and camera position | The selected performance and requested framing |
| Feet and floor | Foot contact, sliding, shadows and reflections | Visible floor and camera height in the scene reference |
| Light and colour | Subject brightness and light direction match the setting | Scene lighting and how much relighting is acceptable |
| Old scene remnants | Unwanted desks, screens, text or lights anywhere in the clip | Specific removal instructions and the complete result |
These are review steps, not measured causes of failure or guaranteed fixes. For a strict requirement to preserve the original performer’s visible footage, assess a foreground-compositing workflow rather than treating a similar generated character as unchanged footage.
Your own footage, sound and cost
The BAD background option uses its prepared sample. For your own clip, choose a workflow that accepts the video and the intended scene edit. Recast’s upper person and face swap tool handles a different task and does not have the same BAD background controls. This fixed-camera example does not establish results for long scenes, moving cameras or multiple shots.
Templates do not include the original song. Add your own recording or music you are allowed to use after reviewing the visuals. Browsing is free; video generation requires a paid account and credits. If you generate a scene image first, inspect that quote separately from the video quote. Resolution, model and preparation choices affect what the workspace quotes; use those current values.
Choosing face-only or whole-person replacement
Following an existing dance with your own character
Sources and example records
The office and rooftop cat examples were recorded on 2026-10-09. The rooftop run used the 15-second office human sample, a prepared character reference and a scene reference, with Wan 3 at 720p and portrait 9:16 framing. The workflow record confirms completion and a charge matching the displayed quote; it does not measure motion accuracy. This article reuses the files in a muted comparison and commissions no new generation. The performer’s private reference photo is not published.
- Adobe Express: removing a video background
- Runway: a generative backdrop workflow
- Alibaba Cloud: video generation and editing task types
External documentation explains the method choices. FlyAIgh’s current workspace determines its available inputs, limits and prices. The three articles in this Recast series reuse some footage for different questions; they are not three independent experiments or evidence of an average success rate.
Frequently Asked Questions
How can I change a video background without changing the person?
If you need the original performer’s image, start with background removal and compositing: isolate the foreground and place it over the new scene, checking the mask edges. Generative editing can reconstruct the performer as well as the setting. Our BAD example is reference-guided generation, not proof of an unchanged original person.
Do I need a green screen to change a video background with AI?
Not every workflow requires one. Background removal tools can isolate a subject without a green screen, while generative workflows use the existing clip as a reference. Neither route guarantees perfect results for every shot. Check moving hands, hair, occlusion and floor contact.
Can I describe a new background, or must I upload an image?
The BAD template offers both. Upload a scene image, or describe a setting and generate a reference image first. Inspect that image before generating the video. Making the scene image and generating the video are separate stages with their own displayed quotes.
What if my background image already contains a person?
The BAD prompt asks the model to use that image for the scene only. The dancer reference specifies the character. Check the prepared image and the output for an unwanted extra person or conflicting appearance; that instruction is not a guarantee that every figure will be ignored.
Can I upload any video to the BAD background template?
BAD uses its prepared office sample as the video reference. For your own footage, choose a workflow that accepts your video and the intended scene edit. Recast’s own-video person and face swap tool has a different purpose; it does not offer the same BAD background controls.
Will the dance moves and camera stay exactly the same?
Exact poses, timing and camera preservation are not guaranteed by reference generation. Compare the same timestamps for turns, gestures, foot contact and the ending. Our example uses a fixed-camera passage; it does not establish the same result for moving cameras or cuts.
The new background looks good, but the character floats or the light looks wrong. What should I check?
Check the scene image’s floor, horizon, camera height and light direction, then compare the character’s feet and shadows throughout the clip. Review the character reference and requested scale too. Change one input or instruction at a time. These are checks, not a diagnosis or a guaranteed repair.
Does changing the background include the original song, and is it free?
Recast templates do not include the original song. Add your own recording or music you are allowed to use after checking the visuals. Video generation requires a paid account and credits. If you generate a scene image first, review that quote separately from the video quote; use the current settings shown in the workspace.