nodes/1038lab-ComfyUI-FlashVSR
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
8 Commits 1 Branch 0 Tags
699140cd31b5a68d5b75b3524f35a62f64dec026
T
Clone
Open with VS Code Open with VSCodium Open with Intellij IDEA
Download ZIP Download TAR.GZ Download BUNDLE
AI Lab 699140cd31 Create __init__.py
2025-10-25 12:38:32 -07:00
.github/workflows
Add GitHub Actions workflow for publishing to registry
2025-10-24 11:14:21 -07:00
FlashVSR
Create __init__.py
2025-10-25 12:38:32 -07:00
LICENSE
Initial commit
2025-10-24 11:10:11 -07:00
pyproject.toml
Add files via upload
2025-10-24 11:31:29 -07:00
README.md
Revise README for ComfyUI-FlashVSR with updates
2025-10-24 17:50:19 -07:00
requirements.txt
Add files via upload
2025-10-24 17:52:11 -07:00

README.md

ComfyUI-FlashVSR ⚡

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

News & Updates

  • 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.

Features

  • High-Quality Video Upscaling: Utilizes the advanced FlashVSR model to upscale videos to 2x or 4x resolution.
  • Multiple Model Versions:
    • Full (Best Quality): High-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.
  • 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.

Installation

Method 1: Install via ComfyUI Manager (Recommended)

  1. Start ComfyUI.
  2. Click the "Manager" button in the sidebar → "Install Custom Nodes".
  3. Search for ComfyUI-FlashVSR.
  4. Click the "Install" button.
  5. Restart ComfyUI.

Method 2: Clone the Repository

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

Method 3: Install via Comfy CLI

  1. Ensure that comfy-cli is installed with:

    pip install comfy-cli

  2. Install ComfyUI-FlashVSR using:

    comfy node install ComfyUI-FlashVSR

  3. 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_3B_FlashVSR_fp32.safetensors Main Diffusion 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

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.

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:

    • safetensors
    • huggingface_hub
    • einops

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

pip install safetensors huggingface-hub einops

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

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

Reference in New Issue View Git Blame Copy Permalink
S
Description
No description provided
Readme GPL-3.0
1.1 MiB
Languages
Python 100%