OpenImagingLab / FlashVSR_Plus

⚡ FlashVSR+ 
 Ruchang Yao 1,2   Shi Guo 1,†   Junhao Zhuang 3   Mingde Yao 2,4   Xin Cai 1,2 
 Jiahao Zhan 2   Bin Fu 1   Xinyuan Chen 1   Yu Qiao 1   Tianfan Xue 1,2,4,† 
 1 Shanghai AI Lab   2 MMLAB CUHK   3 Tsinghua University   4 CPII under InnoHK 
 † Corresponding authors.
 Installation ·
 Model files ·
 Run inference ·
 Python API 
**Your star means a lot for us to develop this project!** :star:
## Overview
**FlashVSR+** is an efficient one-step diffusion framework for high-quality video super-resolution, enabling **spatially untiled 4K inference on a single 24GB RTX 4090 GPU**.
FlashVSR+ introduces two key innovations:
- **Hybrid Attention:** Exploits spatial and temporal attention locality while retaining sparse global interactions, achieving up to **2× DiT acceleration** and **45% lower peak memory**.
- **Reward-Yield Distribution Matching Distillation (RY-DMD):** Enables preference alignment with black-box rewards without differentiating through the reward model, improving perceptual quality and generalizing to text-to-image generation.
We release the inference code, pretrained models, and training dataset to facilitate further research and applications.
## Highlights
- **Three presets:** run Turbo, Base, or V1 through the same CLI and Python API. V1 uses the same inference configuration as Base with its own checkpoints.
- **Large-motion support:** lower `--temporal-compression` from `4` (default) to `2` or `1` to better preserve motion details in videos with large motion.
- **Long-video option:** a numeric `--decode-every` reads and writes video in segments; the default `auto` mode loads the full input.
- **Attention backends:** use Block Sparse Attention, the included Triton operator, or PyTorch SDPA.
## Requirements
- Linux
- NVIDIA GPU with CUDA
- Conda/Miniconda or Python `venv`
- Python 3.10
- Input height and width must both be divisible by 128
The released inference pipeline is CUDA-only and does not support CPU inference.
## Installation
### 1. Create the environment
#### Option A: Conda (recommended)
 [code] 
The editable installation exposes the `flashvsr_plus` Python package and keeps
the `python -m inference.run_stream` command available.
`--no-deps` is intentional. It prevents the package installation from modifying an existing PyTorch/CUDA environment.
#### Option B: Python `venv`
The project also works in a standard Python 3.10 virtual environment. It uses
the host NVIDIA driver/CUDA runtime.
 [code] 
After either setup, activate the environment before installing an optional
attention backend or running inference.
### 2. Install an attention backend
FlashVSR+ supports two optimized sparse-attention backends:
1. **Block Sparse Attention (BSA)** from MIT Han Lab — recommended for NVIDIA data-center GPUs such as **A100, H100, and H200**, where the optimized CUDA operator is supported.
2. **FlashVSR Triton Attention** — our custom Triton implementation, compatible with a wide range of NVIDIA GPUs, including **RTX 4090, A100,