Provenance and Metadata for ML Restoration

Future work. This document outlines an approach to ethical training data documentation for ML-based film restoration. Implementation details are still being developed.

Motivation

ML restoration should be transparent and reproducible. Declaring what was trained, on what sources, with what parameters allows future practitioners and archives to understand and audit the work.

Core Principles

• Declare ML processing — tag outputs with tool, method, and processing date.
• Track provenance — record source/reference element types, scan details, and authorization.
• Record training metadata — dataset shots/frames, steps, model size, checkpoint.
• Classify recovery type — Color (Reference/Non-Reference), Spatial, or Combined.

Proposed Approach

IPTC Digital Source Type

Tag outputs with trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia to communicate that ML processing was involved.

EXR Metadata (Nuke)

Use a ModifyMetaData node before Write to embed provenance keys that downstream tools (Resolve, archival systems) can read:

metadata = nuke.nodes.ModifyMetaData()
metadata['metadata'].fromScript("""
  set exr/MLProcessing "CopyCat Color Recovery"
  set exr/RecoveryType "Reference-Based Color"
  set exr/SourceElement "16mm positive print"
  set exr/ReferenceSource "PAL DVD 2003"
  set exr/ModelCheckpoint "checkpoint_60000"
  set exr/IPTCDigitalSourceType "trainedAlgorithmicMedia"
""")

Sidecar Files

Machine-readable JSON and human-readable text alongside renders. See the previous version of this document for example templates.

Resolve Integration

Import EXR metadata via Resolve’s Metadata panel. Map to Description, Scene/Shot/Take, Keywords, Creator, Rights, Dates. Use Batch Change for common fields.

Standards Context

• IPTC fields communicate AI/ML classification. Pair with institutional policy text.
• C2PA / Content Credentials can be layered later for cryptographic provenance.