HaineiFRDM: Structure-Preserving Diffusion for Film Restoration under Fast Motion and Diverse Defects

HaineiFRDM is a new diffusion-based AI model for film restoration that addresses critical limitations in handling fast motion and complex defects while maintaining structural integrity. The research introduces a patch-wise restoration strategy with frequency-based modules and releases a new film restoration dataset, enabling high-resolution processing on consumer-grade hardware.

HaineiFRDM represents a meaningful advancement in computational film restoration, tackling persistent technical challenges that have plagued existing methods. The core innovation addresses two fundamental problems: the inability to preserve scene structure during fast-motion sequences and inadequate handling of spatially-persistent mixed defects. These limitations have historically resulted in visible artifacts like limb disappearance and structural distortion that degrade restoration quality. The researchers developed a specialized approach leveraging diffusion models' content modeling capabilities to perform content-aware restoration while maintaining visual coherence.

The technical architecture demonstrates engineering sophistication through its patch-wise processing strategy combined with position-aware global fusion modules that preserve cross-patch consistency. The integration of frequency-based enhancement modules specifically targets texture consistency, a known weakness in patch-based approaches. This technical stack enables the model to operate within a 24GB VRAM constraint, democratizing access to professional-grade restoration capabilities beyond specialized studios with enterprise hardware.

The research team's creation of a categorized film restoration dataset with professionally restored reference films and realistic synthetic degradations establishes a new benchmark for the field. This dataset contribution extends the work's impact beyond the specific model, providing infrastructure for future research and development. The commitment to releasing both code and dataset accelerates adoption and reproducibility within the computer vision community.

For practitioners in film restoration, visual effects, and digital archival, this work signals a shift toward AI-driven restoration becoming more accessible and reliable. The ability to handle structural preservation during motion-heavy sequences particularly benefits restoration of older cinema, sports footage, and archival materials where motion blur and defects are prevalent.

• →HaineiFRDM solves fast-motion restoration failures by combining diffusion models with structure-preserving techniques that prevent limb disappearance and distortion artifacts.
• →A novel patch-wise strategy with global fusion modules enables high-resolution film restoration on standard 24GB GPU hardware, reducing professional infrastructure requirements.
• →The released film restoration dataset featuring categorized defects and professionally restored references establishes a new benchmark for evaluating restoration quality.
• →Frequency-based texture enhancement modules and patch-consistent inference frameworks address blocking artifacts inherent to patch-based processing approaches.
• →Open-source release of code and dataset accelerates adoption among researchers and practitioners in film restoration, archival digitization, and visual effects communities.