Chroma keying solved background removal for people who could plan their shoot around it. Everyone else — the vlogger in a rented apartment, the streamer with no room for a backdrop — has been stuck compositing the hard way or not at all. AI matting removes the constraint. Skrrol's Smart Cutout runs a segmentation model on your device, finds the most prominent subject in the shot — a person, a pet, a product, a vehicle — and produces a per-frame alpha matte that behaves exactly like a green-screen key. You can remove a video background in your browser from footage that was never meant to be keyed. This tutorial covers what the AI actually does, when chroma key still wins, the exact steps in Skrrol, and the edge and compositing craft that makes a cutout read as one shot, not a sticker.
What AI matting is and why it works on any footage
A chroma keyer is a color mathematician: it measures how close each pixel sits to a sampled hue and turns everything inside that range transparent. It has no concept of a person — it only knows green, which is why it demands a uniform colored backdrop. AI matting comes at the problem from the opposite direction. A segmentation model is trained to recognize subjects, so instead of asking which pixels are green, it asks which pixels belong to the subject. The background can be a cluttered kitchen, a park, or the back seat of a moving car — the model is not keying on color at all.
In Skrrol this ships as Smart Cutout. Select a clip, open the AI panel, run it, and the model processes the footage locally, returning a per-frame alpha mask — for every frame, a map of which pixels are subject and which are background. From there the workflow is identical to any keyed composite: masked pixels go transparent, the track below shows through, and you refine the edge until the boundary reads as natural. Nothing about the shoot had to be planned for keying, which is why this is the tool for footage that already exists — phone clips, webcam recordings, drone shots, archive material, generated clips born in a model rather than a camera.
AI matting vs chroma keying: different questions, different failures
Because the two tools ask different questions, they fail in different ways, and the failure modes are how you pick between them. A chroma key fails when the color signal degrades: the screen was lit unevenly, green spilled onto the subject's shoulders, somebody wore a green jacket. An AI matte fails when recognition turns ambiguous: two people overlap and the model must decide who the subject is, a dark sleeve sits against an equally dark doorway, fast motion smears the boundary. Neither failure is fatal — both tools carry refinement controls — but they are fixed differently: color problems with tolerance and spill suppression, recognition problems with edge refinement, temporal smoothing, and, for the stubborn frame, a brush.
When chroma key is still the better tool
If you have a properly shot green screen clip, use the chroma keyer — that is not nostalgia, it is the right engineering call. A uniform backdrop turns background removal into a clean color problem, and Skrrol's keyer is built for exactly that: an eyedropper to sample the key color, tolerance, edge feather, and choke sliders that update in real time while the timeline plays, spill suppression that neutralizes the green or blue cast on skin and hair without desaturating the rest of the subject, and a matte view that shows the alpha in black and white so holes are visible before export. The eyedropper samples any color — blue screens, magenta backdrops, custom painted walls — and the keyer is tuned for consumer realities: uneven lighting, slightly wrinkled cloth, hair detail, motion blur.
The decision rule is control. Chroma key when you controlled the shoot; Smart Cutout when you did not. A recording corner you use every week — a streaming setup, a course-recording desk — justifies a green backdrop and buys you real-time keying with precise spill control. Footage that already exists, was shot out in the world, or came out of a generator calls for the AI matte. The two coexist in one project: a talking-head shot on green gets keyed while a phone-shot insert goes through Smart Cutout, and both composite identically on the timeline.
Step-by-step: remove a video background in Skrrol
Here is the whole pipeline, from import to export.
- Import your clip and drop it on a video track. Any source works — phone footage, webcam, drone, archive — with no special shooting setup.
- Place the new background on the track below: a still image, an AI-generated plate, or another video clip.
- Select the foreground clip, open the AI panel, and click Smart Cutout. The model runs locally and produces a per-frame alpha mask — most short clips finish in seconds on a modern laptop.
- Refine the matte edge: raise edge softness slightly for hair and fur, or raise choke to tighten hard product edges.
- Fix stray frames with the manual brush — paint pixels back in or out on a keyframe, and the correction interpolates across neighboring frames automatically.
- Judge the live composite, then export — the finished composite to MP4, or the cutout alone as a transparent WebM with its alpha channel intact.
One habit worth stealing from compositors: get the real background in place before you refine the matte. Edge decisions are relative — a softness value that looks perfect against a transparency checkerboard can fall apart against a bright sky, and a chewed hairline invisible over a busy night street glares over a clean white sweep. Refine against what the audience will actually see.
Edge quality: hair, motion blur, and contrast
Hair is where a matte earns or loses its credibility. Real hair is not a hard boundary — strands are thinner than a pixel, and the edge of a head is a gradient of partial transparency. Clamp that gradient to a hard line and the subject reads as a paper cutout. The countermeasure is edge softness: raise it slightly and the matte boundary feathers, so wisps and fur strands survive instead of being crunched into a cardboard outline. Treat the slider as a scalpel, not a hammer — a small amount restores the natural gradient; a heavy hand puts the whole silhouette out of focus.
Motion is the second test. A fast-moving hand is not a crisp hand — it is a smear that is partly subject and partly background, and the matte has to preserve that in-between transparency for the motion to look natural. Frame-by-frame AI masks carry a specific risk here: a boundary that lands slightly differently on consecutive frames shimmers on playback. Skrrol runs temporal smoothing across the mask sequence to damp that flicker and hold the edge steady. If a sizzle survives on a tricky passage, that is a brush job, not a slider job.
Contrast is the lever you control at the shoot. The model traces a boundary most cleanly when the subject visually separates from what is behind it — a charcoal sweater against a charcoal sofa asks it to find an edge that barely exists. If you know footage is headed for matting, three easy wins: keep some distance from the background, avoid dressing the subject in the background's tones, and light the subject slightly brighter than the room. For hard-edged subjects — products, bottles, boxes — there are no wisps to preserve, so tighten with choke instead of softening and strip the faint rim of leftover background from the silhouette.
The brush: fixing frames and picking between subjects
No automatic matte is perfect on every frame, which is what the manual brush is for. Switch to brush mode, scrub to the offending frame, and paint — additively to restore pixels the model dropped, like a hand that briefly fell out of the mask, or subtractively to remove pixels it grabbed, like a chair back mistaken for a shoulder. The correction does not stay trapped on that frame: brush edits interpolate to neighboring frames automatically, so fixing one keyframe usually heals the surrounding stretch. Run the automatic pass, watch the clip once at full speed, and spot-fix the moments that jump out — painting every frame by hand defeats the point of the tool.
Multi-subject shots have a deliberate default: Smart Cutout isolates the dominant subject in the frame. Two people in the shot, one mask. When you want both — or want them handled separately, say to layer them at different scales — duplicate the clip, run the cutout on each copy, and use brush corrections to steer each pass toward a different person. Two copies, two mattes, independent control over every subject in the frame.
Compositing the cutout over a new background
A clean matte is half the job; the other half is convincing the eye that subject and background were photographed together. Three matches do most of that work. Match the light: if the plate's key light comes from the left and your subject was lit from the right, flip the composition or pick a different background — no grade hides a light-direction conflict. Match the color temperature: a subject shot under cool household LEDs dropped onto a golden-hour field needs warming before the frame reads as one photograph. Match the focus: blur the background slightly if it sits behind the subject, because needle-sharp on needle-sharp flattens the shot into a collage.
The composite renders live on the timeline — subject above, background below, no separate render pass — so iteration is cheap: swap the plate three times, scrub the motion, judge each version at speed. This is also where the AI-native workflow gets interesting: generate a scene with a model like Sora, drop it under your cutout, and put a real person inside a location that never existed. Export the final composite to MP4, or — if compositing happens downstream in another tool — export the cutout as a transparent WebM with alpha.
Privacy: the model runs on your device
The typical online background remover uploads your video to a server, runs the model there, and sends the result back. Smart Cutout does not work that way: the segmentation model runs on your device, the clip is processed locally in the browser, and the source video never leaves it. No upload, no processing queue, no copy of your footage on somebody else's infrastructure. If you cut client work under NDA, footage of your kids, or an unreleased product, that is not a nicety — it decides whether you may use the tool at all.
The same posture holds on the chroma side — keying runs locally in the browser on GPU shaders, and source clips never leave your device — so whichever removal path a clip takes, your media does not travel. There is a practical bonus, too: with no upload leg, processing starts the moment you click, so the seconds you spend are matting seconds, not transfer seconds.
Where this lands: face-cams, ad variants, generated scenes
Gaming creators sit at the center of this workflow. The cut-out face-cam floating over gameplay is the genre's signature layout, and it used to require a green screen behind the chair. With AI matting the webcam recording goes straight in: run Smart Cutout, scale the subject into a corner, and composite over the gameplay capture. Streamers who keep a permanent green screen should keep keying it — real-time keying and spill suppression are exactly for that — but a highlight recorded without the backdrop is no longer condemned to the rectangular webcam box.
Ad and social teams get a different win: variant speed. Cut a presenter or product out once, then swap the background per placement — one plate behind the vertical TikTok cut, another behind the YouTube pre-roll, a seasonal variant for the holiday flight — without reshooting a frame. The subject stays pixel-identical across versions, which keeps brand review painless while the background carries the creative rotation; pair the cutout with generated plates and one shoot becomes a campaign's worth of settings.
Start small. Take one clip you already have — a webcam take, a phone shot of a product, ten seconds of your dog in the yard — drop a background under it, and run Smart Cutout. Refine the edge, brush out one bad frame, export. The loop takes minutes, and once it clicks you stop thinking of the background you shot as the background you are stuck with.