Color encyclopedia · AI Magic Mask 2
Magic Mask 2 is Resolve's AI-assisted isolation tool for selecting a subject directly in the Viewer, tracking it through the shot and turning the result into a usable matte for local grading. This page focuses on the real workflow and controls visible in Resolve 21.
What Magic Mask actually does
The value of Magic Mask is not that AI grades the shot for you. It removes a large part of the manual isolation work. You still decide what to correct, how strong the grade should be and whether the mask is clean enough for the shot.
Start from a clear point in the Viewer so Resolve understands which subject you want to isolate.
Overall Track propagates the selection through time instead of making you rebuild it frame by frame.
Use edge, noise and black/white cleanup only where the generated mask actually needs help.
Once the matte is stable, use the node for exposure, colour, saturation or background separation.
Real Resolve 21 panels
The screenshots below show Magic Mask in the actual Color workspace, after the first click and after the overall tracking pass.
Magic Mask 2 lives directly inside the normal node-based Color workflow. You create the isolation in the Viewer and use it as the basis for local grading — for example to adjust a person, object or background without building a complex HSL key by hand.
After the first click, Resolve creates a separate isolation item and shows it on the Magic Mask timeline. The right side contains quality and matte-refinement controls, so you can improve the mask instead of repeatedly trying to select the subject from scratch.
After tracking, the blue Overall Track bar shows the calculated range and the green check confirms the completed pass. At this stage, the important job is to inspect the difficult parts of the motion rather than assuming AI guarantees a perfect matte on every frame.
Matte refinement
The right-side controls solve different problems: edge geometry, noise, softness and black/white cleanup. Small deliberate adjustments usually look better than one aggressive setting.
Faster is useful for the first pass and for testing the idea. Better makes more sense once the subject is selected correctly and final mask quality matters more than processing speed.
Smart Refine helps improve difficult selection edges. It does not replace checking the Viewer; it is an extra refinement stage for hair, fine details and complex boundaries.
This group controls geometric matte refinement. Shrink or the opposite expansion changes the boundary, Shape defines the refinement behaviour, Radius controls the affected width and Iterations controls the number of passes.
Use these controls with restraint. Smoothing stabilises a rough edge, Denoise removes small mask noise and Blur Radius softens the boundary. Strong values can quickly create an artificial halo.
This redistributes edge softness inward or outward relative to the mask boundary. It is useful when the matte is mostly correct but either eats into the subject or spills too far into the background.
These are matte-cleanup tools: remove grey contamination from the black area, strengthen the white subject region and make the mask more decisive. Use them for final cleanup rather than trying to rescue a bad initial selection.
A final filtering stage after the main matte refinement. It is best used once the mask structure is already correct and only small artefacts remain.
Practical workflow
The order matters because a strong grade amplifies every weakness in the mask.
Magic Mask is easier to control when the local task lives in its own node. This keeps AI isolation, the local grade and its intensity separate from the primary balance.
Start from a clear part of the subject rather than a difficult semi-transparent edge. A strong initial point saves time during later matte refinement.
After Overall Track, inspect fast motion, occlusions, hair and scale changes. These are the places where drift and rough edges are most likely to appear.
The stronger the grade, the more obvious a weak edge becomes. Refine the mask first, then change exposure, colour, saturation or the background.
Magic Mask vs classic tools
AI is powerful, but it is not automatically the best answer. Qualifier, Windows and Magic Mask each have their own sweet spot.
Best when the target is defined by a stable colour or luminance range: skin, sky, a product colour or a highlight band.
Best when geometry is enough: face vignette, background darkening, product emphasis or a controlled region of the frame.
Best when you need the subject itself and its shape changes through the shot: people, moving objects and complex silhouettes.
FAQ
Short practical answers about tracking, matte refinement and the difference from classic secondary tools.
No. Qualifier isolates pixels mainly by colour and luminance ranges. Magic Mask 2 uses AI-assisted subject isolation from viewer clicks and tracks those references through time.
Add and Subtract clicks change the AI interpretation and can be tracked through the shot. Paint is a frame-by-frame patch tool for short mask errors; it does not replace a better click-based selection when the problem lasts for many frames.
It stores the clicks used to guide mask generation and shows where each reference has been tracked. The Overall Track row makes gaps in tracking easy to spot before final refinement.
Faster generates a lower-quality mask more quickly and is useful for setup or garbage matting. Better produces more detail and costs more processing time, so it is the logical choice for the final visible matte.
Hair, translucent fabric, fast motion and occlusion are common low-confidence areas. First improve the clicks and tracking, then use Consistency and Matte Finesse controls such as Smoothing, Denoise and Blur only as much as necessary.
Resolve 21 can render the Magic Mask into cached frames and create a separate external matte node. That makes the generated alpha reusable elsewhere in the node graph instead of keeping the AI mask only inside its original node.
AI Magic Mask 2 in Resolve 20 moved from the older stroke-centric workflow to additive/subtractive clicks with separate paint corrections and Matte Finesse. Its primary purpose is fast secondary-grading selection of a person, object or feature, not replacing full roto/compositing. Resolve 21 can Render in Place a tracked Magic Mask into a separate external matte node, which is useful once a heavy AI track is approved and should become a stable reusable matte.
| Stage | Tool | Purpose | What not to do |
|---|---|---|---|
| Initial selection | Add click | tell the AI which object/region to include | placing dozens of clicks before the first review |
| Exclusion | Subtract click | identify a similar object or region that should be excluded | repairing a systemic mistake with hundreds of paint strokes |
| Tracking | Overall / individual click tracks | propagate the selection through time | tracking the whole clip without inspecting difficult occlusions |
| Local repair | Plus / Minus Paint | fix specific frames/regions where the AI failed | using Paint as manual roto for the whole shot |
| Matte edge | Refine Range, Denoise, Blur, In/Out Ratio, Clean Black/White | fit the edge to the type of grade | creating a hard matte for a soft colour correction without reason |
| Complex scene | Window / Qualifier + Magic Mask | constrain the AI spatially or by signal | expecting one AI mask to solve transparency, hair detail and occlusion perfectly |
| Approved track | Resolve 21 Render in Place Magic Mask | create a cached external matte node for downstream work | rendering the matte before the track and edges are actually approved |
Place the first click on a frame where the subject shape is readable and motion blur/occlusion are not at their worst. A clearer initial hint usually means less manual repair later.
Add a new click only when it explains a new problem to the AI: similar background, another part of the subject or a false inclusion. Do not clutter the Click List before evaluating the first track.
Use Paint tools and individual-click retracking locally where the error actually appears. After a repair, inspect frames before and after it so the matte does not pop.
A soft relight/skin grade can tolerate a more forgiving edge; a strong hue/exposure separation needs a cleaner matte. Matte Finesse should serve the correction rather than the desire for a pretty black-and-white mask.
the AI lacks enough discriminating information or the scene contains similar shapes/textures
a subtract click on the false subject, a spatial Window and retracking the problem section
fine semi-transparent details change faster than the mask model can hold them consistently
Refine Range, Denoise/Blur, a softer grade and an additional Window/Qualifier
matte finesse is making a semi-transparent edge too binary
reduce Clean/Clip, restore Blur/In-Out Ratio and judge the final grade rather than the matte alone
the external matte was created before the track was fully verified
return to the original Magic Mask, fix the track and render the matte again
The Magic Mask menu adds Render in Place Magic Mask. Resolve caches the tracked mask as frames and creates a separate external matte node connected to the working node. This reduces processing load during further grading and lets the approved matte behave like a conventional external alpha source.
