Skip to content

Ultra fast frame interpolation using Rife Tensorrt inside ComfyUI

License

Notifications You must be signed in to change notification settings

huchukato/ComfyUI-RIFE-TensorRT-Auto

 
 

Repository files navigation

ComfyUI Rife TensorRT ⚡

python cuda trt by-nc-sa/4.0

node

This project provides a TensorRT implementation of RIFE for ultra fast frame interpolation inside ComfyUI

This project is licensed under CC BY-NC-SA, everyone is FREE to access, use, modify and redistribute with the same license.

If you like the project, please give me a star! ⭐


⏱️ Performance

Note: The following results were benchmarked on FP16 engines inside ComfyUI, using 2000 frames consisting of 2 alternating similar frames, averaged 2-3 times

Device Rife Engine Resolution Multiplier FPS
H100 rife49_ensemble_True_scale_1_sim 512 x 512 2 45
H100 rife49_ensemble_True_scale_1_sim 512 x 512 4 57
H100 rife49_ensemble_True_scale_1_sim 1280 x 1280 2 21

🚀 Installation

Navigate to the ComfyUI /custom_nodes directory

git clone https://github.com/huchukato/ComfyUI-RIFE-TensorRT-Auto
cd ./ComfyUI-RIFE-TensorRT-Auto
pip install -r requirements.txt

⚠️ CUDA Version Selection

This node defaults to CUDA 13 (RTX 50 series, driver 580+).

For CUDA 12 (RTX 30/40 series):

pip install -r requirements_cu12.txt

For CUDA 13 (Default):

pip install -r requirements.txt

📦 CUDA Toolkit Required

The node automatically detects your CUDA installation via CUDA_PATH or CUDA_HOME environment variables.

If CUDA is not detected, download from: https://developer.nvidia.com/cuda-13-0-2-download-archive

🎯 Resolution Limits

Important: The TensorRT engine supports different resolution ranges based on the selected profile:

  • small profile: 384-1080px
  • medium profile: 672-1312px
  • large profile: 720-1920px (perfect for 1440x960 and higher resolutions)

For images larger than your selected profile's maximum, resize them before using the RIFE node or select a higher profile.

🎯 Resolution Profiles

The node supports resolution profiles to optimize VRAM usage:

  • small: 384-1080px (recommended for most video generation)
  • medium: 672-1312px (for higher resolution videos)
  • large: 720-1920px (for 4K and high-resolution content)
  • custom: Connect a "RIFE Custom Resolution Config" node for manual control

The following RIFE models are supported and will be automatically downloaded and built:

  • rife49_ensemble_True_scale_1_sim (default) - Latest and most accurate
  • rife48_ensemble_True_scale_1_sim - Good balance of speed and quality
  • rife47_ensemble_True_scale_1_sim - Fastest option

Models are automatically downloaded from HuggingFace and TensorRT engines are built on first use.

☀️ Usage

  1. Load Model: Insert Right Click -> Add Node -> tensorrt -> Load Rife Tensorrt Model

    • Choose your preferred RIFE model (rife47, rife48, or rife49)
    • Select precision (fp16 recommended for speed, fp32 for maximum accuracy)
    • Select resolution profile (small, medium, or custom)
    • The model will be automatically downloaded and TensorRT engine built on first use
  2. Process Frames: Insert Right Click -> Add Node -> tensorrt -> Rife Tensorrt

    • Connect the loaded model from step 1
    • Input your video frames
    • Configure interpolation settings (multiplier, CUDA graph, etc.)

🤖 Environment tested

  • Windows 11, CUDA 13.0, TensorRT 10.15.1.29, Python 3.12, RTX 5090
  • WSL Ubuntu 24.04.03 LTS, CUDA 12.9, TensorRT 10.13.3.9, Python 3.12.11, RTX 5080

🚨 Updates

February 2026

  • Fixed Dependencies: Updated TensorRT to 10.15.1.29 to resolve installation conflicts
  • RTX 5090 Support: Tested and confirmed compatibility with RTX 5090
  • Resolution Documentation: Added clear guidance on resolution limits and preprocessing

January 2026

  • CUDA 13 Default: Updated to CUDA 13.0 and TensorRT 10.14.1.48
  • Auto CUDA Detection: Automatically finds CUDA toolkit and DLL paths
  • Resolution Profiles: Added small/medium/custom profiles to reduce VRAM usage

December 2025

  • Automatic Model Management: No more manual downloads! Models are automatically downloaded from HuggingFace and TensorRT engines are built on demand
  • Improved Workflow: New two-node system with Load Rife Tensorrt Model + Rife Tensorrt for better organization
  • Updated Dependencies: TensorRT updated to 10.13.3.9 for better performance and compatibility

👏 Credits

License

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)

About

Ultra fast frame interpolation using Rife Tensorrt inside ComfyUI

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%