Glossary

Terms specific to the chroma and spatial recovery workflows in this repository. Standard Nuke node names (Read, Write, Merge, Transform, Crop, etc.) are not redefined here — consult Foundry documentation for those.

ACES 2065-1

Linear AP0 scene-referred color space for archival/exchange masters. Recommended delivery space with EXR half.

Alignment

Pixel-accurate registration of Reference to Source, including crop, so branches differ only in the intended characteristic (color or spatial).

Chroma (Cb, Cr)

Color-difference channels in YCbCr. In chroma recovery, these come from the Reference branch.

CopyCat

Nuke node that trains a compact CNN using supervised Input/Target pairs assembled in the node graph.

F_Align

Furnace alignment operator used as an automated first pass before manual Transform refinement.

Film gauge

Physical film width (16mm, 35mm). Different gauges have distinct spatial characteristics and damage patterns.

Film generation

Duplication stage (original negative, internegative, print, duplicate). Generational differences affect spatial quality.

Ground truth (Target)

The desired output used as Target during training:

Chroma recovery: Source Y + Reference Cb/Cr
Spatial recovery: Reference Y + Source Cb/Cr

Hold-out frames

Frame pairs excluded from training, used only for validation.

Inference

Nuke node that applies a trained .cat model to new frames/sequences.

Internegative

Intermediate negative created from a positive element. Often higher spatial quality than distribution prints.

Luma (Y)

Brightness channel in YCbCr. In chroma recovery, luma comes from the Source branch.

MatchGrade

Nuke node for look matching. Used in these workflows as an optional baseline to compare against learned reconstruction.

Provenance

Recorded history of sources, processes, and parameters enabling reproducibility and auditability.

Telecine

Video transfer of a film element from earlier preservation workflows. May preserve cleaner sections despite lower resolution.

Training pair

A single supervised example: Input (degraded) and Target (ground truth) sharing identical content and alignment.

YCbCr

Color model separating luma (Y) from chroma (Cb, Cr). Used to construct ground truth by recombining channels from Source and Reference.