StereoAlign

StereoAlign [SA] belongs to Fusion's Studio-only Stereo Nodes. It corrects the relationship between the left and right views after a valid stereo pair and disparity data have already been prepared upstream.

The essential model is disparity, not generic optical flow. A pair of ordinary RGBA images without the required disparity information does not provide the data StereoAlign uses for its disparity-dependent interpolation and warping.

A common separate-eye chain is Left/Right sources -> Disparity -> StereoAlign. If depth or disparity is required later, add another Disparity node after StereoAlign and calculate it again on the corrected pair.

That rebuild is mandatory for a technical reason: StereoAlign changes RGBA geometry and consumes/destroys the incoming auxiliary disparity channels. The old disparity values no longer describe the new pixel coordinates.

In an unstacked separate-eye topology StereoAlign has two physical image inputs, Left Input and Right Input, and two physical image outputs, Left Output and Right Output. Reducing it to one Image Input and one Image Output hides the actual stereo signal flow.

Stack Mode makes the topology dynamic. The captured Resolve 21 Inspector shows None, Horiz, Vert and Layers. In packed horizontal, vertical or layered workflows the whole stereo representation can travel through the primary Left path, so a duplicated Right cable should not be fabricated.

Blackmagic documentation also uses the term Separate for the unstacked state, while the captured Resolve 21 UI used by this article displays None. The article keeps the captured label and explains its meaning instead of pretending all versions use identical wording.

Vertical Alignment, Convergence Point and Eye Separation live in one node so several stereo geometry corrections can be performed within a single resampling stage instead of stacking unnecessary transforms.

Vertical Alignment should normally be diagnosed first. Even small vertical disparity is tiring for the viewer and can make later convergence decisions harder to judge.

Vertical Alignment Apply to selects which eye receives the correction. Mode selects Global or Per Pixel behavior: Global is a uniform shift, while Per Pixel is intended for disparity-aware local correction where supported by the stereo data.

Y Shift is the global vertical pixel offset. Snap to Whole Pixels can avoid an unnecessary fractional resample when the measured error is truly integral, but it can be too coarse for a real subpixel mismatch.

The vertical Sample action should be evaluated against the incoming stereo/disparity geometry. Do not treat a changed StereoAlign result as if it still shared the upstream disparity coordinate system.

Convergence Point changes the horizontal relationship and therefore the zero-parallax plane. X Shift moves the views in X, while Apply to chooses Left, Right or Split.

Split shares the convergence correction between the eyes and therefore reduces the absolute translation applied to either view. Left or Right keeps the opposite eye as a master/reference.

Convergence is not Eye Separation. Convergence moves the zero-parallax plane; Separation scales stereo disparity/depth strength. Diagnose these operations independently.

Eye Separation Apply to selects Left, Right or Both, while Separation is a scale-like adjustment with 0 as the neutral baseline. Large changes can expose regions that were never visible in either source eye.

StereoAlign cannot reconstruct information that does not exist in either view. Strong separation changes or warps may expose holes or transparent gaps; this can be a source-coverage limitation rather than a simple control error.

Each eye has its own Depth Ordering. Largest Disparity On Top and Smallest Disparity On Top decide which disparity region wins where warped surfaces overlap. Judge the option on a real foreground/background occlusion boundary.

Clamp Edges is intended for small frame-boundary gaps. It stretches edge pixels, so it is not a restoration tool for large missing regions. The documented conditional Edge Softness reduces the visibility of that stretch.

Source Frame and Warp Direction can combine up to four contributors: Left Forward, Right Forward, Left Backward and Right Backward. They differ both by source frame and disparity warp direction.

Left Forward means left source with forward disparity; Right Forward means right source with forward disparity; Left Backward means left source with back disparity; Right Backward means right source with back disparity. These are reconstruction contributors, not four quality levels.

Additional contributors can fill gaps because the other eye may contain useful texture, but conflicting disparities at an occlusion can also create doubled contours. Enable contributors one at a time and keep only those that solve a specific boundary.

The captured Refine Edges section proves the Enable control. Do not invent Strength, Radius or HiQ as mandatory current controls without target-build evidence; conditional subcontrols should be confirmed separately.

Stack Mode defines representation: None, Horiz, Vert or Layers in the captured target UI. Horizontal and vertical modes pack both eyes into one raster layout; Layers carries a layered stereo representation. Upstream and downstream topology must agree with this mode.

Swap Eyes exchanges left and right order. It fixes a reversed stereo pair, but it does not repair bad disparity, vertical mismatch or incorrect convergence.

Quality control must include playback, near/far objects and occlusion boundaries. Inspect fine edges at 100% with high-quality viewing because low-quality or nearest-neighbor preview can hide or exaggerate interpolation artifacts.

Lens distortion or other camera geometry differences that interfere with matching are best normalized before disparity calculation. StereoAlign is an alignment stage, not a replacement for camera calibration or undistortion.

A robust production order is: validate Left/Right plus disparity, fix vertical alignment, set convergence, adjust eye separation, resolve occlusion/edge behavior, then choose Stack Mode and eye order. Rebuild disparity afterward if later nodes need it.