workflow-templates-broken-zap-2026-06-08/site/knowledge/models/flashvsr.md

FlashVSR

FlashVSR is a diffusion-based streaming video super-resolution framework that achieves near real-time 4× upscaling through one-step inference with locality-constrained sparse attention.

Model Variants

FlashVSR v1

• Initial release of the one-step streaming VSR model
• Built on Wan2.1 1.3B video diffusion backbone
• 4× super-resolution optimized

FlashVSR v1.1

• Enhanced stability and fidelity over v1
• Improved artifact handling across different aspect ratios
• Recommended for production use

Key Features

• One-step diffusion inference (no multi-step denoising required)
• Streaming architecture with KV cache for sequential frame processing
• Locality-Constrained Sparse Attention (LCSA) prevents artifacts at high resolutions
• Tiny Conditional Decoder (TC Decoder) achieves 7× faster decoding than standard WanVAE
• Three-stage distillation pipeline from multi-step to single-step inference
• Runs at ~17 FPS for 768×1408 videos on a single A100 GPU
• Up to 12× speedup over prior one-step diffusion VSR models
• Scales reliably to ultra-high resolutions

Hardware Requirements

• Minimum: 24GB VRAM (A100 or similar recommended)
• Optimized for NVIDIA A100 GPUs
• Significant VRAM required for high-resolution video processing
• Multi-GPU inference not required but beneficial for throughput

Common Use Cases

• Real-world video upscaling to 4K
• AI-generated video enhancement and artifact removal
• Long video super-resolution with temporal consistency
• Streaming video quality improvement
• Restoring compressed or low-resolution video footage

Key Parameters

• scale: Upscaling factor (4× recommended for best results)
• tile_size: Spatial tiling for memory management (0 = auto)
• input_resolution: Source video resolution (outputs 4× larger)
• model_version: v1 or v1.1 checkpoint selection