Real-Time Diffusion-Based Streaming Video Super-Resolution
Upscale and sharpen low-resolution videos in near real time. FlashVSR-v1.1 is the first diffusion-based one-step streaming video super-resolution framework, delivering state-of-the-art quality at up to ~17 FPS on a single A100 GPU with v1.1 adding enhanced stability and fidelity over the original release.
What it does
FlashVSR-v1.1 is a diffusion-based video super-resolution (VSR) model that takes low-resolution video and reconstructs it at significantly higher resolution (optimized for 4× upscaling). What sets FlashVSR apart from other diffusion-based VSR tools is its speed: instead of running many denoising passes per frame, it produces results in a single step and processes frames in a streaming fashion. This makes diffusion-quality video super-resolution practical at scale, even at ultra-high output resolutions.
Problem it solves
- Low-resolution video sources – Upscales legacy, web-compressed, or low-bitrate footage to sharp, high-resolution output (4× recommended)
- Slow diffusion VSR pipelines – Most diffusion-based super-resolution tools are far too slow for practical use on video. FlashVSR completes super-resolution in a single inference step, achieving up to ~12× speedup over prior one-step diffusion VSR models
- Streaming / real-time use cases – Built as a streaming framework, enabling near real-time processing rather than batch-only workflows
- Ultra-high resolutions – Scales reliably to very large output resolutions where competing tools tend to break down or introduce artifacts
- Stability across aspect ratios and content types – v1.1 specifically targets improved stability and fidelity over the initial v1 release
Input/Output
- Input:
- A low-resolution video (optimized for 4× super-resolution input)
- Output: A high-resolution MP4 video file
- Sharp, detailed reconstruction with preserved fine structures and textures
- Temporally coherent motion across frames, processed in a streaming pass
Accuracy & Speed
- Speed & Scale
Processing speed | ~17 FPS for 768 × 1408 video on a single NVIDIA A100 GPU |
Speedup vs. prior work | Up to ~12× faster than previous one-step diffusion VSR models |
Infrastructure needed | Single GPU; optimized for NVIDIA A100 / A800 (Ampere). Runs on H200 (Hopper) with reduced acceleration. Other GPUs (RTX 40/50, H800) currently unverified |
Output quality | State-of-the-art among diffusion-based VSR models |
Technical Details
Architecture | Diffusion-based one-step streaming VSR framework |
Approach | Three-stage distillation pipeline that enables streaming super-resolution in a single inference pass |
Attention | Locality-Constrained Sparse Attention (LCSA) — cuts redundant computation and bridges the train–test resolution gap |
Decoder | Tiny Conditional Decoder for fast, high-quality reconstruction |
Training Data | VSR-120K — purpose-built dataset of 120k videos and 180k images |
Backend | Requires Block-Sparse Attention library for full LCSA acceleration |
Recommended setting | 4× video super-resolution |
Developer | Junhao Zhuang et al. (Tsinghua University / OpenImagingLab) |
Model Variants
FlashVSR is available in two released versions:
Version | Best For |
FlashVSR v1 | Original release (Oct 2025) |
FlashVSR v1.1 | Recommended — enhanced stability and fidelity (Nov 2025) |
Each version also ships full and tiny inference variants, including a long-video tiny variant for extended sequences.
Compliance & Provenance
Provider | Open-source |
Provider type | Specialized |
License | Apache 2.0 |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary | Pending — provider has not yet published per Art. 53(d) |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- HuggingFace: https://huggingface.co/JunhaoZhuang/FlashVSR-v1.1
- License: apache-2.0