★ Best Paper Award, 2025
FACTNet
A Flash-Attention-optimized CNN–Transformer framework for real-time video restoration, recognised with a Best Paper Award.
Overview
FACTNet is a hybrid CNN–Transformer architecture designed for real-time video restoration. It couples the local feature extraction strengths of convolutional networks with the long-range modelling of Transformers, while a Flash-Attention-optimized attention mechanism keeps inference fast enough for real-time use.
Highlights
- Flash-Attention optimization reduces the memory and latency cost of self-attention, enabling real-time throughput.
- A CNN–Transformer hybrid design balances fine-grained detail recovery with global temporal consistency.
- Recognised with a Best Paper Award at its 2025 venue.
Why it matters
Real-time restoration unlocks practical deployment in streaming, surveillance, and low-quality video enhancement where latency budgets are tight. FACTNet shows that attention-heavy restoration can be made efficient without sacrificing quality.