Video
Active
SwitchX
No green screen, no manual masking, and no separate relighting pass tacked on afterward.
SwitchX is a video-to-video compositing model built by Beeble.
A lot of video work runs into the same wall: the footage itself is fine, but the setting is wrong. The shot was filmed in a spare room that needs to look like a studio. The lighting doesn't match the rest of the campaign. Or the same take has to work across five markets with five different backdrops, and reshooting five versions isn't realistic.
SwitchX works out where the subject sits in every frame, lifts it out at the pixel level, places it onto the new background, and adjusts the lighting on the subject so it looks like it was actually filmed there. Hair, motion blur, soft edges, and shadows are preserved rather than rebuilt, which is what keeps the result from looking like a cutout pasted on top of a photo.
Three parts of the model do most of the work, and they're worth understanding before you build around it.
alpha_mode: autoMost compositing setups need either a green screen or someone tracing a mask by hand. With alpha_mode set to auto, SwitchX looks at the footage and works out what the subject is — a person, a product, an animal, whatever's in frame — and generates the alpha matte itself, including the fine detail around hair and semi-transparent edges that's normally the hardest part to get right. Footage shot in an ordinary room under ordinary light works fine; there's nothing special to rig on set.
Generating one accurate mask is one thing. Keeping it accurate across hundreds of frames as the subject moves, turns, or is briefly blocked by something passing through the shot is the part that normally eats up an editor's afternoon. SwitchX defines the subject mask on the first frame and then propagates it through the rest of the clip, tracking the subject frame to frame so the mask holds up through motion and partial occlusion without being redrawn by hand.
Swapping the background is the easy half. The harder half is that the subject was lit for the old scene, and against a new one it'll look pasted-in unless the lighting agrees too. SwitchX reads the lighting cues in the reference image — color temperature, direction, intensity — and adjusts how the subject is lit in the output so it sits in the new scene the way it would have if the camera had actually been rolling there.
Billed per 30 output frames
Billed per generated image
A few situations where this kind of compositing tends to come up.
Same talent, same take, a different backdrop for every region, retailer, or seasonal campaign — without booking the talent again.
Move a product shot from a plain set into a lifestyle scene, with lighting adjusted to match the new setting.
Extend or replace a set without building it physically or renting an LED volume for the day.
Swap out backgrounds containing text, branding, or location-specific details so the same shot reads correctly in every market.
Give older or stock footage a new setting that fits a current campaign's look, instead of shelving it.
A lot of video work runs into the same wall: the footage itself is fine, but the setting is wrong. The shot was filmed in a spare room that needs to look like a studio. The lighting doesn't match the rest of the campaign. Or the same take has to work across five markets with five different backdrops, and reshooting five versions isn't realistic.
SwitchX works out where the subject sits in every frame, lifts it out at the pixel level, places it onto the new background, and adjusts the lighting on the subject so it looks like it was actually filmed there. Hair, motion blur, soft edges, and shadows are preserved rather than rebuilt, which is what keeps the result from looking like a cutout pasted on top of a photo.
Three parts of the model do most of the work, and they're worth understanding before you build around it.
alpha_mode: autoMost compositing setups need either a green screen or someone tracing a mask by hand. With alpha_mode set to auto, SwitchX looks at the footage and works out what the subject is — a person, a product, an animal, whatever's in frame — and generates the alpha matte itself, including the fine detail around hair and semi-transparent edges that's normally the hardest part to get right. Footage shot in an ordinary room under ordinary light works fine; there's nothing special to rig on set.
Generating one accurate mask is one thing. Keeping it accurate across hundreds of frames as the subject moves, turns, or is briefly blocked by something passing through the shot is the part that normally eats up an editor's afternoon. SwitchX defines the subject mask on the first frame and then propagates it through the rest of the clip, tracking the subject frame to frame so the mask holds up through motion and partial occlusion without being redrawn by hand.
Swapping the background is the easy half. The harder half is that the subject was lit for the old scene, and against a new one it'll look pasted-in unless the lighting agrees too. SwitchX reads the lighting cues in the reference image — color temperature, direction, intensity — and adjusts how the subject is lit in the output so it sits in the new scene the way it would have if the camera had actually been rolling there.
Billed per 30 output frames
Billed per generated image
A few situations where this kind of compositing tends to come up.
Same talent, same take, a different backdrop for every region, retailer, or seasonal campaign — without booking the talent again.
Move a product shot from a plain set into a lifestyle scene, with lighting adjusted to match the new setting.
Extend or replace a set without building it physically or renting an LED volume for the day.
Swap out backgrounds containing text, branding, or location-specific details so the same shot reads correctly in every market.
Give older or stock footage a new setting that fits a current campaign's look, instead of shelving it.
.webp)
.png){kind=small}
.png){kind=small}
