ComfyUI Upscaler TensorRT ⚡

This project provides a Tensorrt implementation for fast image upscaling inside ComfyUI (2-4x faster)

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

For commercial purposes, please contact me directly at [email protected]

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

---

⏱️ Performance

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

🚀 Installation

• Install via the manager
• Or, navigate to the /ComfyUI/custom_nodes directory

git clone https://github.com/yuvraj108c/ComfyUI-Upscaler-Tensorrt.git
cd ./ComfyUI-Upscaler-Tensorrt
pip install -r requirements.txt

🛠️ Supported Models

• 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).

  • 4x-AnimeSharp
  • 4x-UltraSharp
  • 4x-WTP-UDS-Esrgan
  • 4x_NMKD-Siax_200k
  • 4x_RealisticRescaler_100000_G
  • 4x_foolhardy_Remacri
  • RealESRGAN_x4
  • 4xNomos2_otf_esrgan

☀️ Usage

• 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_to to resize the upscaled images to fixed resolutions

🔧 Custom Models

• 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 as shown:

  ComfyUI-Upscaler-Tensorrt/__init__.py

  Line 77 in 8f7ef5d

  "model": (["4x-AnimeSharp", "4x-UltraSharp", "4x-WTP-UDS-Esrgan", "4x_NMKD-Siax_200k", "4x_RealisticRescaler_100000_G", "4x_foolhardy_Remacri", "RealESRGAN_x4"], {"default": "4x-UltraSharp", "tooltip": "These models have been tested with tensorrt"}),

• 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

🚨 Updates

4 March 2025 (breaking)

• 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

⚠️ Known issues

• 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 .pth to .onnx

🤖 Environment tested

• Ubuntu 22.04 LTS, Cuda 12.4, Tensorrt 10.8, Python 3.10, H100 GPU
• Windows 11

👏 Credits

• NVIDIA/Stable-Diffusion-WebUI-TensorRT
• comfyanonymous/ComfyUI

License

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