Video & AI

Video & AI tools

Cutout rotoscoping with SAM-powered Smart Select, Generative Erase, the AI Depth Mask and a full Keyer — all on-device, never metered.

Four effects that used to need a VFX department. All of them run on your Mac — the models ship with the app or download once (SAM 2 is a single ~94 MB fetch, shared by every project), nothing is uploaded, nothing is billed per use, and everything works offline.

Cutout — rotoscoping

Separates a subject from its background with a soft, per-pixel matte — hair-and-fur quality edges, not a traced outline. On a still: Generate Cutout and it’s done; when it finds several subjects, chips let you keep one, several, or all (Pick in viewport to click the one you mean).

Smart Select is the precision tool: click or sweep over anything — a person, a shirt, a face — and the model finds its edges from the image itself. The gestures compose the way you’d hope: a Keep click asks “what object is here” generously; an Alt-click removes a whole thing; an Alt-sweep carves away only what the sweep covers, shaped by the image — the way to peel an attached shadow off a face without losing the face. Two refine brushes finish the edge: Hard Edge snaps a wavering outline crisply onto the picture’s own boundary, and Hair re-reads the picture where you paint — strands stay covered, the milky haze between them clears.

On footage it tracks. Press ▶ and the matte carries through the clip — identity locked (a passerby detected mid-shot never joins your subject), edges steadied, your Smart Select clicks riding along on optical flow so the mask re-finds the image every frame instead of decaying. The playhead follows so you watch it hold; pause the moment something looks wrong (nothing is lost — pause is pause, not cancel), fix that frame, and track again from it — ◀ carries a perfected middle frame back toward the start. Hand corrections are remembered as pixels that ride the footage: a stray hair painted off a moving head stays off. Corrected frames become anchors that a later re-track respects instead of overwriting.

Finish with Clean Up (removes stray disconnected blobs — whole blobs only, it can never thin hair), Refine Edge, Grow/Shrink, Feather, Invert — and three views: Result, Matte (white-on-black — the classic check for holes), and Overlay, which keeps everything removed on screen, dimmed red, so there’s something to aim at when adding it back. Exports always render the Result. Bake to Alpha MOV writes the isolated layer as ProRes 4444 with a real alpha channel for other software.

Generative Erase

Remove a thing; the hole fills from its surroundings. The model is bundled (MIT-licensed, so it ships inside the app): no download, no account, no cost per use.

On a still: make a selection → Generative Erase in Tool options. All the fills compute up front (~a sixth of a second), so the Variation row switches instantly between genuinely different machines: Auto picks by how deep the hole is; Texture copies real patches from around the selection (it cannot invent — cardboard comes back as cardboard); Smooth interpolates from the edge (right for walls, skies, paper); Detail / Detail 2 let the model synthesize structure (right over foliage, brick, hair — and over a plain wall it will confidently invent a window, which is exactly why the choice is yours). To new layer puts the fill on its own layer, so the original keeps every pixel.

On footage: Mark the object with the same Smart Select gesture, then Erase Across Clip. The fill is synthesized once and carried by motion tracking — re-asking the model only where the picture demands it (revealed background, accumulated drift) — which is what keeps it rock-steady instead of shimmering. Watch it hold, pause where it breaks, re-mark, sweep on. Show Patches tints every pixel the fill owns — the check before trusting it.

Depth Mask

AI depth estimation turned into a grading tool: the model runs once per still (about 39 ms), and everything after is realtime dial-turning — this scrubs like an ordinary effect where other apps’ depth tools crawl.

From and To select a distance band (0 = nearest thing in frame, 100 = farthest — “From 50, To 100” is the background, which is why those are the defaults), with Feather softening the band in depth and Soften/Grow shaping the resulting matte in pixels. The band carries a built-in grade — Brightness (±2 stops), Saturation (0 = the B&W-background shot), Blur (fake shallow depth of field, blurred through the band’s own weight so the sharp subject never smears into its own halo) — and it lands visibly, a mild push-the-background-back, so the dials always show what they do. Keyframe From and To together and the grade travels: a focus pull with two keyframes.

The views are deliverables, not just diagnostics: Overlay (tints everything outside the band — the one to dial against), Depth (the map itself, exportable as a pass), Matte (track-matte anything against it), and Cut Out (the band becomes the layer’s alpha — drop the subject straight onto a new background, no matte dance). And the sleeper feature — Limit Effects Below: every effect under this one in the stack is confined to the band. Add Depth Mask, switch it on, add a Gaussian Blur underneath: only the background blurs, on one layer.

Where the model guesses wrong (a mirror reading as the room beyond it), the Correct Depth brushes overrule it: Nearer / Farther / Set To, with Snap to Edges keeping a rough stroke on the surface you started it on — a hand held in front of the mirror keeps its own depth. On footage, Track Depth Across Clip estimates every frame and holds the range steady (each frame refitted to the last — the cure for AI depth’s characteristic flicker), corrections ride along on flow, and a clip with a camera solve pins the whole shot into one absolute depth range: nothing breathes, nothing drifts.

Five presets ship: Blur Background, B&W Background, Darken Background, Spotlight Subject, Dreamy Distance.

Keyer

Five genuinely different ways to decide what’s background, sharing one professional matte chain:

• Chroma — colour-difference keying against a screen colour: the green/blue-screen workhorse, with Gain, Balance (for screens shot warm or cool), Pre-blur (cleans the 4:2:0 chroma blocking in the matte only — the image is never blurred), and real Despill with its own balance.
• Luma — matte from brightness; Color — distance to a flat colour (the cel-animation mode — works on greys, where hue-reasoning fails), with Also Key for extra colours; Difference — against a clean plate (park on a frame with the subject out of shot → Use frame); Unmult — elements shot on black (fire, smoke, sparks) composited correctly with colour divided back out.
• The shared tail: Clip Black (screen dirt), Clip White (holes in the subject), Matte Gamma, Shrink/Grow (choke), Feather (matte-only), Invert — and Protect: colours the key must hold onto, forced back solid. The answer to a key eating dark hair or a blue shirt, which no amount of stacking keyers can express.

Nothing keys until you eyedrop the screen — so the original pixels stay there to sample. Views: Result, Matte, and Status — Keylight’s three-tone check (black/grey/white) that finds holes and screen dirt invisible against a busy background. Exports always render Result, whatever view you parked in. Masks on the layer act as garbage mattes — draw one to cut rigging out rather than fighting Clip Black.

Recipe — subject pop on any photo. Depth Mask → preset B&W Background → nudge From until the subject’s whole depth stays coloured. One effect, the classic isolation grade.

Recipe — cel-clean key on flat art. Keyer → Mode Color, eyedrop the paper. Tolerance low, Softness ~10 — flat fills key exactly where a chroma keyer would overthink them.