This project provides a Tensorrt implementation for fast image upscaling using models inside ComfyUI (2-4x faster)
Last tested: 12 January 2026 (ComfyUI v0.8.2@c623804 | Torch 2.9.1 | Tensorrt 10.14.1.48 | Python 3.12.3 | RTX5090 | CUDA 13.1 | Ubuntu 24.04)
If you like my projects and wish to see updates and new features, please consider supporting me. It helps a lot!
Note: The following results were benchmarked on FP16 engines inside ComfyUI, using 100 identical frames
| Device | Model | Input Resolution (WxH) | Output Resolution (WxH) | FPS |
|---|---|---|---|---|
| RTX5090 | 4x-UltraSharp | 512 x 512 | 2048 x 2048 | 12.7 |
| RTX5090 | 4x-UltraSharp | 1280 x 1280 | 5120 x 5120 | 2.0 |
| RTX4090 | 4x-UltraSharp | 512 x 512 | 2048 x 2048 | 6.7 |
| RTX4090 | 4x-UltraSharp | 1280 x 1280 | 5120 x 5120 | 1.1 |
| RTX3060 | 4x-UltraSharp | 512 x 512 | 2048 x 2048 | 2.2 |
| RTX3060 | 4x-UltraSharp | 1280 x 1280 | 5120 x 5120 | 0.35 |
This node features fully automatic CUDA detection and TensorRT installation! When ComfyUI loads the node for the first time, it will:
- Auto-detect your CUDA version (12 or 13)
- Install the appropriate TensorRT packages automatically
- Configure everything for seamless operation No manual steps required! Just clone the repo and restart ComfyUI.
If you prefer manual installation or encounter issues: Auto-install scripts:
# Linux/macOS
./install.sh
# Windows
install.bat
# Python (cross-platform)
python install.pyManual requirements files:
# For CUDA 13 (RTX 50 series)
pip install -r requirements.txt
# For CUDA 12 (RTX 30/40 series) - LEGACY METHOD
pip install -r requirements_cu12.txt💡 Note: The requirements_cu12.txt is provided as a legacy fallback method. The automatic installation is strongly recommended as it handles CUDA detection and package installation seamlessly.
The node automatically detects your CUDA installation via CUDA_PATH or CUDA_HOME environment variables.
CUDA_PATH
CUDA_HOME
If CUDA is not detected, download from: https://developer.nvidia.com/cuda-13-0-2-download-archive
-
These upscaler models have been tested to work with Tensorrt. Onnx are available here
-
The exported tensorrt models support dynamic image resolutions from 256x256 to 1280x1280 px (e.g 960x540, 512x512, 1280x720 etc).
- Load example workflow
- Choose the appropriate model from the dropdown
- The tensorrt engine will be built automatically
- Load an image of resolution between 256-1280px
- Set
resize_toto resize the upscaled images to fixed or custom resolutions
- To export other ESRGAN models, you'll have to build the onnx model first, using export_onnx.py
- Place the onnx model in
/ComfyUI/models/onnx/YOUR_MODEL.onnx - Then, add your model to this list load_upscaler_config.json
- Finally, run the same workflow and choose your model
- If you've tested another working tensorrt model, let me know to add it officially to this node
- Add more resizing scale factors
- Add custom resolution resizing
- Support 4x-UltraSharpV2_Lite, 4x_UniversalUpscalerV2-Neutral_115000_swaG, 4x-ClearRealityV1
- Load models from config PR#57
- Merge yuvraj108c#48 by @BiiirdPrograms to fix soft-lock by raising an error when input image dimensions unsupported
- Automatic tensorrt engines are built from the workflow itself, to simplify the process for non-technical people
- Separate model loading and tensorrt processing into different nodes
- Optimise post processing
- Update onnx export script
- If you upgrade tensorrt version, you'll have to rebuild the engines
- Only models with ESRGAN architecture are currently working
- High ram usage when exporting
.pthto.onnx
- Ubuntu 24.04, Debian 12
- Windows 11
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)