GlobalAlign

GlobalAlign [GA] is a Studio-only Stereo node that manually applies global X/Y translation and rotation to the left and right eye. It does NOT compute optical flow; use it to correct broad misregistration before downstream disparity estimation, not to repair local lens distortion or per-object stereo mismatch.

The manual describes an orange Left Input that can accept the left-eye image or an already packed stereo image, and a green Right Input for separate right-eye footage. It also documents two outputs, Left Output and Right Output. Availability of right-hand ports depends on the packing mode. No dedicated Effect Mask input is described for GlobalAlign.

Translation X and Translation Y each have a Balance menu and a numerical Shift. None applies no translation. Left moves the LEFT image only; Right moves the RIGHT image only. Split moves both views oppositely. The eye named by the menu is the MOVING eye, not an untouched reference.

Use Y Shift to remove vertical error measured on matching landmarks. X Shift changes the horizontal stereo relationship and can alter perceived screen depth: do not attempt to eliminate all legitimate parallax at different depths. Correct roll and vertical mismatch before a deliberate convergence check.

Each translation axis offers Snap to Nearest Pixel. Full-pixel shifts avoid some interpolation-induced softness, but fractional shifts can be necessary for accurate sensor rectification. Evaluate texture and edges at 100% in both eyes, not just in one fit-to-view preview.

Rotation has its own Balance plus Angle. The manual illustrates a requested 10° roll: with Left or Right, the named eye turns 10°; with Split, each eye turns 5° in opposite directions. Rotation can expose empty corners and cannot remedy local projective discrepancies.

Translation Filter Method governs image reconstruction during alignment. All 11 options shown on the archived dropdown are explained, from Nearest Neighbor through Bessel. No filter has a universally ideal setting: compare fine texture, ringing, haloing and inter-eye sharpness under motion.

Visualization (left eye) includes None, R-C, Y-B, G-M and Diff. These are diagnostic comparison encodings, not a creative grade or a final stereo delivery format. Even correctly aligned views retain real depth parallax, so Diff will not vanish everywhere.

The archived Stack Mode screenshot lists None, Horiz, Vert and Layers, whereas the textual manual calls separate-eye routing Separate. The manual also specifies that separate right-hand connectors are conditional. This naming discrepancy is noted explicitly; the exact mapping of the current UI choices to exposed Flow ports needs native verification.

Swap Eyes reverses the stereo order in applicable packed workflows. It is not a geometric fix. Before exporting, confirm the actual left/right labels, their depth ordering, and whether an Anaglyph preview is a diagnostic only or a required deliverable.

A common processing chain is separate left/right source images into GlobalAlign, then into Disparity or a stereo preview node. The teaching registry uses semantic port IDs LeftInput/RightInput and LeftOutput/RightOutput. They validate illustrative dataflow but are not the exact native scripting connector IDs.

The site’s graphs are reviewed against typed image cables and documented upstream/downstream node contracts. They do not prove the native conditional-port visibility, numeric alignment units, actual pixels, or viewer comfort. Use a Resolve Studio playback and external stereo QC for the final acceptance.