Disparity

Disparity estimates feature correspondence between two views of the same scene. It is available in DaVinci Resolve Studio and Fusion Studio; one ordinary monocular frame is not a stereo pair.

The eye images retain their RGB and alpha data, while the calculated vectors are stored in the Disparity auxiliary channel. Viewing or exporting only RGBA does not guarantee that this field has been preserved.

The field has horizontal (X) and vertical (Y) components expressed as pixel shifts rather than normalized coordinates or metric distances. X is generally most informative for stereo depth; Y helps reveal vertical misalignment.

Direction matters: Left Output contains left-to-right correspondence, while Right Output carries right-to-left correspondence. Near occlusions these fields need not be exact negatives of one another.

Start with Stack Mode. The captured Inspector offers None, Horiz, Vert, and Layers. None means two separate eyes; earlier Blackmagic manuals call this wiring mode Separate. None does not disable stereo analysis.

For separate images, route the left view into the orange Left Input and the right view into the green Right Input. Both output paths are used in this configuration; the frames should be synchronized.

For Horiz (side by side), Vert (top and bottom), or Layers, both views live in one packed image. Feed that image to Left Input; the separate Right input and output are not needed and may be hidden in packed modes.

Swap Eyes changes the interpretation of eye order. Check the actual source layout before toggling it: the switch does not increase depth or repair a noisy solve caused by mismatched images.

Prepare the sources before tuning the solver. Match exposure, white balance, and conspicuous grade differences: correspondence uses visible colors and gradients, so photometric differences create false matches.

Crop invalid or black borders after transforms. If there is a large global vertical offset, correct that roughly before Disparity rather than wasting refinement passes on a simple alignment error.

Decide where lens distortion is removed. A disparity field calculated in one geometry may no longer describe the pixels after a substantial downstream warp.

Method offers Classic and Advanced in the captured Controls tab. These are visible choices, but the labels alone do not prove a particular undocumented solver implementation. Hold the shot and method fixed during comparisons.

In Advanced, Warp Count controls the number of refinement passes. Extra passes can help a difficult shot, but once correspondence converges they mostly add computation; compare the same problematic object edge.

Iteration Count controls iterative solver refinement. Additional iterations can improve convergence, but cannot invent matches within genuine occlusions or repair reversed eye order.

Smoothness regularizes local noise. Too much smoothing erases thin detail and softens depth discontinuities around hair, foliage, and foreground silhouettes.

Half Resolution trades analysis detail for setup speed. Disable it for final quality control, then inspect at 100% Viewer scale and in playback.

The captured Advanced state is Warp Count 5, Iteration Count 10, Smoothness 10.0, and Half Resolution Off. These are values from one saved node state, not recommended defaults for every clip.

Inspect X and Y disparity separately. A substantial coherent Y shift suggests revisiting upstream vertical alignment; broken X detail around edges suggests matching or occlusion problems.

Do not expect a perfect field over specular reflections, flat areas, repeating patterns, or occlusions. Some pixels visible in one eye simply do not exist in the other; maximizing refinement settings cannot resolve that ambiguity.

The Disparity node does not support RoI or DoD as a bounded-computation shortcut. Plan performance around source preparation, sensible Advanced settings, draft Half Resolution, and caching instead.

Inspect playback as well as stills. Temporal shimmer often appears on moving boundaries and highlights. If spatial correspondence is sound, SmoothMotion can be considered separately for temporal stabilization of the AUX field.

StereoAlign consumes the disparity field while changing stereo geometry; its output no longer contains the original usable disparity. Add another Disparity after the alignment if a later node needs fresh correspondence.

DisparityToZ performs a separate conversion from stereo displacement to Z using suitable camera and stereo-geometry assumptions. Pixel shifts alone must not be described as measured physical depth.

For heavy projects, cache the expensive calculation in EXR or another compatible intermediate only after checking that Disparity AUX survives both writing and reading. An RGBA-only render cannot reconstruct the field.