This presentation explores experimental machine learning techniques in film restoration, focusing on the development of small custom trained AI models tailored to the needs of archival materials. Unlike commercial AI tools optimized for contemporary media, these models are designed specifically to address the unique forms of degradation found in historical film elements. By working with localized datasets and film specific characteristics, the approach avoids overgeneralization and preserves the distinct aesthetics of the original material.
The presentation covers restoration tasks such as color recovery, either guided by reference materials (such as prints, internegatives, or digitized analog elements) or inferred from culturally or artistically analogous sources when references are unavailable, and spatial repair techniques including gauge alignment, generational recovery, and analog video reconstruction.
Emphasis is placed on ethical considerations, particularly the use of locally executed models trained only on authorized data, thereby respecting rights and provenance while ensuring archival transparency. This work argues for a shift toward practical and ethically sourced AI tools that empower archives to perform restoration work at scale without compromising historical integrity or legal clarity.
