Shared Workflow — Stages 0–2

This page covers everything both recovery modes share: Resolve export, Nuke project setup, dataset curation, alignment, shared crop, and the branch decision. Follow this page first, then continue in the guide for your chosen branch.

  • Chroma Recovery — when detail is intact but color is faded, collapsed, or shifted.
  • Spatial Recovery — when color is acceptable but detail, sharpness, or grain are weaker than the reference.

Workflow overview End-to-end overview of the recovery workflow.

Stage 0: Resolve Export + Nuke Project Setup

Source and Reference Preparation

What matters is whether the reference preserves better information for the problem you are solving — not whether it is newer, sharper, or higher resolution.

Raw source before balancing Source that should be technically balanced before training.

Balanced source for training Balanced source plate used as cleaner input to the workflow.

Source preparation:

  • Technically balance the source (neutral, not creative).
  • Remove severe flicker, dirt, splice flashes, and instability that would poison training.
  • For chroma recovery: degrain the source for training if grain interferes with chroma learning. Document settings and keep the original plate.
  • For spatial recovery: preserve original grain structure — do not degrain unless the reference is also degrained.
  • Global cast neutralization: if the source has a strong bias from dye fade or scanning, apply a neutral pre-balance. Recommended: Faded Balancer DCTL/OFX.
Before (raw faded scan) After (Faded Balancer applied)
Raw faded scan Balanced scan

Faded Balancer DCTL neutralizing magenta dye fade in Resolve before Nuke ingest.

Reference preparation:

  • Clean enough to remove transfer artifacts that would mislead the model.
  • For chroma recovery: suppress dust, compression noise, banding (light denoise/deband/median). Geometry changes and temporal warping are not recommended.
  • For spatial recovery: remove dust/dirt/scratches only. Do not median filter or blur the reference — preserve all spatial detail (grain, sharpness, edge definition). For magnetic/video references, target only obvious compression artifacts using tools that preserve spatial frequency content.

Geometry/stabilization: prefer cleanup that does not alter geometry. If you must stabilize or reframe, apply identical transforms to both exports.

Resolve Export

Resolve reference prep Source and reference placed in the same Resolve container.

  1. Conform both sources in a single timeline. Disable retimes, effects, and per-clip grades.
  2. Align the reference to the source in Edit/Inspector (Translate/Scale/Rotate). Allow letterbox/pillarbox; keep stable framing.
  3. Use ACES project settings. Export both with Rec.709 2.4 ODT to EXR — keeps values bounded in [0–1], which CopyCat expects.
  4. Verify parity: resolution, pixel aspect, frame range/rate, channel set (RGB only; omit alpha).
  5. Note any global offsets (scale/translate/rotate) for later reference.

Rules:

  • Same frame range, framing, and resolution.
  • No creative grading.
  • Fix interlacing, cadence problems, and decode issues before export.

QC checklist:

  • Dimensions, frame ranges, and pixel aspect match
  • Channel sets match (RGB), values in 0–1 when read into Nuke
  • No inadvertent retimes or additional color transforms

Nuke Project Setup

Project settings:

  • Color management: OCIO with ACES 1.2 or ACES 1.3.
  • Working space: ACEScg (scene-linear, AP1).
  • Viewer process: ACES ODT matching your display (e.g., ACES 1.0 SDR-video / Rec.709 2.4).

Read node settings (both Source and Reference):

  • Training pairs (from Resolve Rec.709 2.4 ODT): Read.colorspace = Utility - sRGB - Color Picking.
  • ACES masters (interchange/comp): Read.colorspace = ACES - ACES2065-1. If used for training, transform to display-referred (apply Rec.709 2.4 ODT) rather than clamping naively.

Verify:

  • Toggle Viewer between Source/Reference; confirm consistent appearance under the chosen ODT.
  • Confirm identical ingest transforms on both branches.

ACES and Color Management Reference

Training domain (recommended): Display-referred — export Rec.709 2.4 ODT, ingest via Utility - sRGB - Color Picking, process in ACEScg, build YCbCr ground truth with identical chains. Values are naturally bounded; only light safety clamping needed.

Alternative — Log domain: Viable in theory but significantly slower and, in testing, inferior for chroma recovery fidelity. Use only if footage demands it.

Not recommended — Naive linear ACES clamp: Ingesting ACES 2065-1 and clamping to [0–1] crushes highlights and harms both chroma and spatial mapping. If starting from ACES masters, transform to display-referred first.

Both Input and Target must share the exact same domain and transforms. Do not mix linear and display-referred between branches.

Write nodes (delivery):

  • Archival: Write.colorspace = ACES - ACES2065-1 (AP0), EXR 16-bit half (ZIP/DWAA).
  • Review/proxy: Rec.709 ODT → ProRes/H.264. Document viewing intent and ODT.

Resolve interop:

  • ACES 1.2/1.3 project; export ACES 2065-1 EXR masters for Nuke ingest.
  • For display-referred training, export Rec.709 2.4 ODT and ingest with Utility - sRGB - Color Picking.
  • Keep frame size, PAR, and channels identical.

Stage 1: Dataset Curation

Build a small teaching set from representative frames — do not throw the whole sequence into training.

Dataset curation Building paired training examples in Nuke.

Selection Criteria

  • Source frames: intact luma/texture, representative grain, minimal gate weave. Avoid motion-blur-dominated frames unless matched in reference.
  • Reference frames: same shot/timecode when available, or a well-constructed proxy. Avoid heavy compression, baked-in subtitles/logos, unstable grades.
  • Exclude: pairs with occlusions unique to one side (flashes, splice marks) that the model cannot reconcile.

Pair Counts

Scope Pairs Notes
Shot 4–9 Add more if convergence stalls
Scene 12–24  
Sequence 24–64+ Scale with variability

For short ranges (e.g., frames 20–60), anchor at beginning/middle/middle/end.

Coverage

Ensure diversity across:

  • Lighting: warm/cool, day/night, interior/exterior
  • Subjects: skin tones, foliage/sky, fabrics, neutrals
  • Extremes: deep shadows, specular highlights, saturated primaries
  • Textures (spatial recovery): fabric, foliage, skin, edges, smooth gradients

Pairing Rules

  • Temporal: match same frame index/timecode. If off-by-one, prefer the frame with maximal static structure overlap.
  • Spatial: identical resolution/orientation. Overscan/crop must be shared (residual differences handled in Stage 2).
  • Color space: both sides under the same transform (e.g., Rec.709 export) so values remain in 0–1.

Nuke Build

  1. Create a FrameHold per selected index on both Source and Reference branches.
  2. Assemble ordered stacks with AppendClip: one for Source (Input), one for Reference (Target).
  3. Keep a staging AppendClip upstream of the one referenced by downstream PostageStamp nodes for safe reordering.
  4. Verify each pair with viewer wipe or Merge (difference) — judge geometry/alignment only, not color.
  5. Label pairs consistently. Maintain a table of indices/timecodes for traceability.

Documentation

  • Record shot IDs, pair indices, and rationale.
  • Note whether references are direct (telecine/DVD/print) or constructed; cite sources.
  • Flag compromises (compression, residual parallax) for review during training.

Stage 2: Alignment

Pixel-accurate alignment with shared crop so branches differ only in the intended characteristic (color or spatial detail).

Alignment workflow Auto and manual alignment paths inside the template.

Strategy

  1. Single global solve with F_Align using a conservative central ROI. Do not iterate.
  2. Evaluate immediately with Merge (difference). If edges/geometry remain, switch to manual Transform with keyframes.
  3. Keep a Dissolve to compare auto/manual paths quickly.

Merge (difference) alignment check Merge (difference) in the viewer: geometry/edges should be near-black. Visible color differences are expected — only structural misalignment is a problem.

Crop and Subtitle Handling

  • Remove black borders/overscan on both branches — do not train on non-image content.
  • Exclude burned-in subtitles/logos. Where unavoidable, animate a shared crop.
  • Apply the exact same crop to Source and Reference (clone/link) so pixel areas correspond.

Crop node settings Shared crop keeping both branches in the same live picture area.

Nuke Build

  • Compare Reference to Source with Viewer wipe and Merge (difference).
  • Use a Dissolve to switch auto/manual paths. Keyframe per frame (0 = auto, 1 = manual) after inspection.
  • Auto path: F_Align with conservative central ROI. Single global solve (Translate/Scale/Rotate/Perspective). No parameter chasing.
  • Manual path: Transform (translate/scale/rotate) keyed as needed. Judge with Merge (difference).
  • Reference Crop (last step): add Crop on aligned Reference to remove overscan/transient overlays. Keep bypassed while solving; enable as final step. Save this node to clone/link in Stage 3. Do not crop Source here.

Verification Checklist

  • Merge (difference) shows only color/photometric differences; geometry/edges near-black. Do not apply Grade/color correction here.
  • No edge shimmer at borders/corners when toggling Source/Reference.
  • Reference crop removes overscan/mattes without hiding alignment cues.

Troubleshooting: Gate weave/parallax on warped multi-generation references — expect manual Transform keyframing to be time-consuming.

Branch Decision — Stage 3

This is where the workflow diverges. The layout is shared; the branch decision is which channels go into the ground-truth target.

Colorspace node settings Convert both branches to YCbCr before channel recombination.

Shuffle node settings Shuffle node used to build the ground-truth target.

CopyBBox node CopyBBox ensures identical bbox across Source and Ground Truth before connecting to CopyCat.

Branch Ground truth target Continue in
Chroma recovery Source Y + Reference Cb/Cr chroma-recovery.md
Spatial recovery Reference Y + Source Cb/Cr spatial-recovery.md

Rules After the Branch

  • Do not combine chroma and spatial recovery in the same target build.
  • Validate one pass before attempting the second.
  • Train sequence-level first; split to shot-level only where the wide pass fails.
  • Run inference on the full source, not just training frames.
  • Compare against a baseline (MatchGrade) to prove the model is doing something specific.

Common Failure Modes

  • Training on an unstable source (flicker, dirt, severe imbalance).
  • Trusting a bad reference just because it is higher resolution.
  • Keeping misaligned frames in the dataset.
  • Leaving borders, subtitles, or overlays inside the training area.
  • Expecting one model to solve every shot in a difficult sequence.

Back to top

Workflow template for film preservation and restoration research.