ComfyUI-FlashVSR

ComfyUI-FlashVSR
★ 95

视频超分实时处理低显存优化SageAttention加速
ComfyUI-FlashVSR：基于 FlashVSR 的实时视频超分节点，支持 2x/4x 升频、低显存模式与 SageAttention 加速，适合流式处理帧序列。
💡 在 ComfyUI 中对视频帧序列进行实时流式高质量 2x/4x 超分。
🍴 11 Forks💻 Python🔄 2025-11-17
📦
网盘下载
复制链接后前往夸克网盘下载
https://pan.quark.cn/s/c1eafc754fbb
📦 requirements.txt
#
Core
dependencies
torch>=2.0.0
torchvision>=0.15.0
numpy>=1.24.0
einops>=0.6.0
safetensors>=0.4.0
tqdm>=4.65.0
pillow>=9.5.0
huggingface_hub>=0.19.0
#
Optional
performance
optimizations
(not
required)
#
Uncomment
if
you
want
SageAttention
support
for
~20-30%
speedup:
#
sageattention>=1.0.0
#
triton>=2.1.0
📄 README

ComfyUI-FlashVSR ⚡

A powerful ComfyUI custom node based on the FlashVSR model, enabling real-time diffusion-based video super-resolution for streaming applications.

https://github.com/user-attachments/assets/1d1528c5-e3c1-487f-9c29-267ddb817809

Features

High-Quality Video Upscaling: Utilizes the advanced FlashVSR model to upscale videos to 2x or 4x resolution.

  • Multiple Model Versions:
  • Full (Best Quality): Highest quality results with significant VRAM usage.
  • Tiny (Fast): Balanced quality and speed for faster processing.
  • Tiny Long (Low VRAM): Optimized for GPUs with limited VRAM, ideal for long videos.
  • SageAttention Optimization (Optional): Automatic ~20-30% speedup when SageAttention is installed. Falls back gracefully if not available.
  • Intelligent Tiling: Supports enable_tiling to process high-resolution videos efficiently on low-VRAM GPUs.
  • Automatic Model Download: On the first run, the node will automatically download the required .safetensors models from Hugging Face (1038lab/FlashVSR).
  • Audio Passthrough: Maintains the original audio during video frame processing, ensuring synchronization and quality preservation.

    • News & Updates

    2025/11/15: FlashVSR 1.1 Model Update + Frame Duplication Fix ( update.md )

  • Added new model: Wan2_1-T2V-1.1_3B_FlashVSR_fp32.safetensors
  • Improved T2V → VSR quality, stability, details
  • Applied frame duplication fix (Issue #3)
  • Updated UPDATE.md accordingly

    • 2025/10/24: Initial release of ComfyUI-FlashVSR.

  • Added FlashVSR ⚡ and FlashVSR Advanced ⚡ nodes.
  • Implemented automatic model download from Hugging Face (1038lab/FlashVSR).
  • Supports .safetensors models, audio passthrough, and tiling for low VRAM.

    • Installation

    Method 1: Install via ComfyUI Manager (Recommended)

  • Start ComfyUI.
  • Click the “Manager” button in the sidebar → “Install Custom Nodes”.
  • Search for ComfyUI-FlashVSR.
  • Click the “Install” button.
  • Restart ComfyUI.

    • Method 2: Clone the Repository

  • Navigate to your ComfyUI custom_nodes directory.
  • Run:
    •    git clone https://github.com/1038lab/ComfyUI-FlashVSR.git

  • Restart ComfyUI.

    • Method 3: Install via Comfy CLI

  • Ensure that comfy-cli is installed with:

    • “`bash

    pip install comfy-cli

    “`

  • Install ComfyUI-FlashVSR using:

    • “`bash

    comfy node install ComfyUI-FlashVSR

    “`

  • Restart ComfyUI.

    • Method 4: Manually Download the Models

  • The models will be automatically downloaded to ComfyUI/models/FlashVSR/ on the first run.
  • To manually download the models, visit 1038lab/FlashVSR on Hugging Face and download the .safetensors files into the ComfyUI/models/FlashVSR/ folder.

    • | Model File | Purpose |

    |———–|———|

    | Wan2_1-T2V-1.1_3B_FlashVSR_fp32.safetensors | New FlashVSR 1.1 Main Diffusion Model |

    | Wan2_1-T2V-1_3B_FlashVSR_fp32.safetensors | Previous FlashVSR 1.0 Main Model |

    | Wan2.1_VAE.safetensors | Video VAE |

    | Wan2_1_FlashVSR_LQ_proj_model_bf16.safetensors | Low-Quality Projection |

    | Wan2_1_FlashVSR_TCDecoder_fp32.safetensors | Tiny Model Decoder |

    📖 For optional performance optimization (~20-30% speedup), see SageAttention Installation Guide

    Usage

    This node processes image (frame) sequences. For a complete video workflow, combine it with other nodes in ComfyUI.

  • Load: Use a video loader (e.g., VHS – Video Load) to load video frames and audio.
  • Process: Connect the frames and audio to the FlashVSR node.
  • Save: Use a video combiner (e.g., VHS – Video Combine) to combine the output frames and audio into a final upscaled video.

    • FlashVSR Nodes

    Optional Settings 💡 Tips

    | Optional Setting | Description | Tips |

    | —————————- | —————————————————————————————— | ————————————————————————————————– |

    | preset (Simple) | Choose between: Fast (Tiny model), Balanced (Tiny model), High Quality (Full model). | High Quality requires significant VRAM. Consider using the Advanced node if you face OOM errors. |

    | model_version (Advanced) | Options: Tiny (Fast), Tiny Long (Low VRAM), Full (Best Quality). | Full offers the best quality, while Tiny Long is optimized for minimal VRAM. |

    | enable_tiling (Advanced) | Breaks the video into smaller chunks to save VRAM. | Enable this if you encounter OOM errors, especially with the Full model at 4x scale. |

    | speed_optimization | Optimizes for processing speed. Higher values yield faster results. | Default is 2.0. |

    | quality_boost | Boosts quality at the cost of VRAM usage. Higher values yield better results. | Default is 2.0. The Full model can handle 3.0 without crashing. |

    | Input Frames | The video frames to process. | Requires at least 21 frames for initialization. |

    | 4x Upscaling | Optimized for 4x upscaling. | 2x upscaling is supported, but 4x generally provides better results. |

    | sageattention (Advanced) | Enable/Disable SageAttention optimization. | Enabled by default. Provides ~20-30% speedup if sageattention package is installed. |

    About FlashVSR Model

    FlashVSR is a real-time diffusion-based video super-resolution model. It is designed to provide high-quality upscaling, particularly suited for streaming applications. The .safetensors versions are included for enhanced compatibility and security.

    Requirements

  • ComfyUI
  • Python 3.10+
  • Required packages:
  • torch >= 2.0.0
  • torchvision >= 0.15.0
  • safetensors >= 0.4.0
  • huggingface_hub >= 0.19.0
  • einops >= 0.6.0
  • numpy >= 1.24.0
  • tqdm >= 4.65.0
  • pillow >= 9.5.0

  • Optional packages (for performance boost):
  • sageattention >= 1.0.0 – Provides ~20-30% speedup (see Optional Performance Optimization)
  • triton >= 2.1.0 – Required by SageAttention

    • These packages are typically included in ComfyUI environments. If you encounter an import error, run:

    pip install torch>=2.0.0 torchvision>=0.15.0 safetensors>=0.4.0 huggingface-hub>=0.19.0 einops>=0.6.0

    Optional Performance Optimization

    For an automatic ~20-30% performance boost, you can install SageAttention:

    pip install sageattention triton

    Note:

  • SageAttention requires a CUDA-capable GPU and may conflict with some ComfyUI environments.
  • For detailed installation instructions and troubleshooting, see SageAttention Installation Guide.
  • If you encounter issues after installing SageAttention, you can:
  • Disable it in the FlashVSR ⚡ Advanced node by setting sageattention to disable.
  • Or uninstall it: pip uninstall sageattention triton
  • The node will work perfectly fine without SageAttention installed – it will automatically fall back to standard PyTorch attention.

    • Troubleshooting

  • FileNotFoundError: Missing Wan2.1_VAE.safetensors:

  • This error usually occurs when the model download fails or is skipped.
  • Fix: Delete the FlashVSR folder in ComfyUI/models/, then restart ComfyUI to trigger the automatic download again.

  • Out-of-Memory (OOM) Error / CUDAMallocAsyncAllocator.cpp error:

  • Occurs when VRAM is exhausted, especially with the High Quality preset or Full model at 4x scale.
  • Fix: Use the FlashVSR Advanced ⚡ node and enable enable_tiling to reduce VRAM usage.

    • Credits

  • FlashVSR: OpenImagingLab/FlashVSR
  • Original HF Models: JunhaoZhuang/FlashVSR
  • Safetensors Models: 1038lab/FlashVSR
  • Created by: AILab

    • Star History

    If this custom node helps you or if you appreciate the work, please give a ⭐ on this repo! It’s a great encouragement for my efforts!

    License

    GPL-3.0 License