DepthMap

Resolve 21 places Depth Map [DMp] in Matte Nodes. Its only required connection is the yellow 2D Input; a second physical blue Effect Mask input is optional. The tool estimates perceived object distance and creates an Alpha channel for downstream matte workflows.

The key data boundary is that DepthMap does not generate an auxiliary Z channel, a multisample Deep Image stream, or recovered scene geometry. It is an AI/perceptual estimate normalized into a matte/Alpha workflow and should not be mislabeled as a physical Z pass.

Depth Map Preview shows the calculated map as a black-and-white diagnostic image while tuning. When Preview is disabled, the flow returns to the normal image representation and the calculated map is used through Alpha. White matte regions are affected/opaque and black regions unaffected/transparent.

Mode selects Faster or Better. The manual describes Depth Map as computationally intensive: Faster improves adjustment responsiveness while Better is the default/best-result mode for final checking. Resolve 21 additionally advertises up to a six-times Fusion Depth Map speed increase over the previous version, but Faster and Better remain different quality modes.

Starting with Resolve 20, Blackmagic automatically uses AI Depth Map v2 in place of the original model while keeping the interface and general workflow. The map therefore should not be taught as a simple luminance conversion; it is an AI estimate of perceived scene distance.

Platform boundary: AI Depth Map v2 is Studio-only inside DaVinci Resolve. Standalone Fusion Studio is itself a licensed product and also includes the native AI Depth Map tool; Blackmagic has shipped Fusion Studio updates specifically accelerating Depth Map. It should not be presented as an ordinary free-Resolve matte tool even though it appears as a native Fusion node.

Invert reverses transparent and opaque map regions. It changes near/far matte polarity for the task at hand; it does not reverse a physical 3D Z axis.

Resulting Map Adjustment is enabled with Adjust Map Levels. Resolve 21 makes an important distinction: with it disabled the full map range is not clipped by this scaling stage; with it enabled values are clipped/remapped into 0–1 as they will behave in Alpha. Far Limit controls the black end, Near Limit the white end and Gamma redistributes intermediate depth values.

Isolate Specific Depth selects a depth band rather than a rectangular image region. Target Depth 1 is foreground and 0 background; Tolerance widens the range on either side and Softness adds a ramp in and out. Think of this as a depth qualifier, not a physical 3D slice measured in scene units.

Map Finesse operates on the resulting Alpha. Post Processing enables the block, Post-Filter smooths the map relative to image contours, Contract/Expand erodes or dilates boundaries and Blur softens the matte edge. These controls can production-clean a usable map but cannot repair a fundamentally wrong AI interpretation of the scene.

Effect Mask is a separate optional blue input applied after the tool is processed. It gates where DepthMap may contribute; it does not supply depth and it does not change Target Depth. Isolation answers which depth band is selected while Effect Mask answers where that result is spatially allowed.

The current captured Resolve 21/21.1 Inspector also exposes Advanced > Deflicker at 1.0. The current Matte Nodes chapter does not define its exact numeric mapping, so this page records it as a temporal-stability control visible in the target UI while keeping detailed direction/units as a live-verification boundary.

For DepthBlur, topology should follow the actual datatype. Because DepthMap generates Alpha, a clean workflow is beauty into DepthBlur.Input, DepthMap.Output into the green Blur Image, and Blur Channel set to Alpha. Selecting Z merely because the tool is named DepthMap would be incorrect because no metric auxiliary Z buffer is created.

DepthMap is useful for depth-aware blur, local grading, atmosphere masks, foreground/background separation and other matte workflows. Hair, transparency, reflections, motion blur, textureless areas and ambiguous perspective can still fool monocular depth inference, so inspect both the map and temporal behavior rather than trusting the AI automatically.

DepthMap is an AI matte generator based on perceived distance rather than a rendered Z pass; its production result lives in Alpha/matte workflows.

Target Depth is relative to the current shot, not a metric distance, because the model infers near/far from visual cues rather than camera-space geometry.

Preview is diagnostic only; once disabled, downstream work should explicitly inspect/use the generated Alpha.

Faster and Better trade responsiveness for solve quality; Better still requires edge and temporal QC.

Invert changes matte polarity rather than reversing a physical Z axis.

Adjust Map Levels is optional and changes how the solve is remapped/clipped into Alpha-oriented 0–1 behavior.

Far Limit, Near Limit and Gamma should be tuned only after the underlying depth hierarchy is plausible.

Isolation converts the continuous estimate into a depth band; Target Depth chooses the center, Tolerance its width and Softness its ramp.

Postprocessing is cleanup after selection, not a repair for wrong AI inference.

Post-Filter can reduce variation, Contract/Expand shifts boundaries, and Blur softens them; all can damage thin edges if overused.

Advanced Deflicker is a temporal-stability control that must be evaluated in motion on the target build.

Effect Mask is a post-solve spatial gate and does not participate in AI depth inference.

DepthMap Output can drive DepthBlur correctly through Blur Image with Blur Channel=Alpha; selecting Z is wrong unless auxiliary Z was created separately.

Hair, glass, reflections, sky, water, smoke, defocus and motion blur are mandatory failure cases for QC.

Upstream crop, resize, grading and Merge can alter visual cues and therefore change the inferred map.

DepthMap should not be used as a technical measurement of real scene geometry; it is a perceptual compositing estimate.