FIAT/IFTA World Conference 2026

Artificial intelligence has rapidly become an important tool in contemporary film restoration. Machine learning systems now assist with chroma reconstruction, density balancing, spatial inference, defect detection, and reference-guided restoration. Despite these advances, current AI methods still encounter significant limitations when dealing with the physical and statistical characteristics of photochemical film and scanned archival materials. This presentation examines the principal technical boundaries of present-day AI restoration systems and identifies the research directions most likely to overcome them. The first major limitation is the loss of high-frequency information: many generative models operate in compressed latent representations that suppress stochastic micro-signal, producing over-smoothed textures and synthetic noise patterns. The second limitation is temporal consistency: motion picture restoration requires stable behaviour of fine detail across frames, yet current systems often produce flickering textures and motion-dependent smoothing. The third limitation concerns acquisition-specific signal characteristics: film stock response curves, dye fading behaviour, scanner optics, and laboratory processing artifacts vary enormously across archival materials. The presentation then examines emerging approaches that may address these limitations. Pixel-space diffusion offers improved microstructure reconstruction but at extreme computational cost. Wavelet-based diffusion frameworks reorganise image information into frequency bands while preserving spatial localisation. Grain transport - treating film grain not as noise to be removed but as a structured stochastic signal to be probabilistically preserved - represents a further frontier. Drawing on the presenter's ongoing experimental work in machine-learning-assisted film restoration, the presentation argues that the most promising path forward lies in hybrid workflows that combine machine learning inference, classical frequency-domain processing, stochastic grain modelling, and expert human supervision.