Recurrent Super-Resolution

Recurrent super-resolution is the video enhancement approach that propagates hidden features across timesteps so each output frame benefits from long temporal context - it offers strong temporal continuity and efficient memory reuse for long sequences.

What Is Recurrent SR?

• Definition: VSR architecture with temporal state passed from frame to frame during inference.
• State Role: Encodes accumulated historical detail and motion context.
• Direction Options: Forward-only for streaming or bidirectional for offline quality.
• Representative Models: BasicVSR-style pipelines with flow-guided propagation.

Why Recurrent SR Matters

• Long-Range Context: Can integrate evidence across many frames beyond fixed windows.
• Temporal Stability: Recurrent propagation encourages coherent output trajectories.
• Memory Efficiency: Avoids storing full long-window feature sets explicitly.
• Quality Potential: Strong results when alignment and propagation are stable.
• Streaming Suitability: Natural fit for online enhancement workflows.

Recurrent SR Components

Propagation Module:
- Warp previous hidden features into current frame coordinates.
- Combine with current frame features for updated state.

Bidirectional Fusion:
- Offline variants run forward and backward passes.
- Merge both directions for higher quality.

Reconstruction Head:
- Convert propagated features into high-resolution frame output.
- Apply temporal consistency losses during training.

How It Works

Step 1:
- For each frame, align previous state using flow and update recurrent features with current observations.

Step 2:
- Decode enhanced frame from updated state and continue propagation through sequence.

Recurrent super-resolution is a long-context VSR paradigm that leverages persistent temporal memory for coherent high-quality enhancement - careful drift control is essential for stable performance over extended clips.