ESRGAN NCNN Models

Collection of pre-converted ESRGAN/Real-ESRGAN models for use with NCNN Vulkan tools (e.g., realesrgan-ncnn-vulkan). Each model directory typically contains a .param and .bin pair ready for inference.

What’s inside

Model folders in this repo:

• 4x_NMKD-Superscale-Artisoftject_210000_G
• 4x_NMKD-Superscale-SP_178000_G
• 4x_NMKD-YandereNeoXL_200k
• 4x_NMKDSuperscale_Artisoft_120000_G
• 8x_NMKD-Typescale_175k
• realesr-animevideov3-x2
• realesr-animevideov3-x3
• realesr-animevideov3-x4
• realesrgan-x4plus
• realesrgan-x4plus-anime

Each directory should include files like:

model.param
model.bin

Some models may provide multiple variants or different scaling factors; consult the folder name for scale hints (e.g., x2, x3, x4, 8x).

Requirements

• NCNN Vulkan executables, e.g. realesrgan-ncnn-vulkan from the Real-ESRGAN NCNN release
  • Linux users can place the binary in your PATH and ensure Vulkan drivers are installed
  • NVIDIA/AMD/Intel GPU with Vulkan support recommended

Use with Upscayl

You can use these models directly in Upscayl Desktop.

• Download Upscayl: upscayl.org/download

Using Models in Upscayl:

1. Download from releases
2. Extract the downloaded ZIP file
3. Open Upscayl and click the Settings tab
4. Click the "Select Folder" button under the "ADD CUSTOM MODELS" section
5. Select the models folder in the extracted repository (or any model folder you want to add)
6. Go back to the main Upscayl screen and select your custom models

Quick start (Real-ESRGAN NCNN)

Basic single-image upscale using a model directory:

realesrgan-ncnn-vulkan \
  -i input.png \
  -o output.png \
  -s 4 \
  -m ./realesrgan-x4plus

Notes:

• -m should point to the directory that contains .param and .bin.
• -s is the scale factor; match it to the model (e.g., 2, 3, 4, 8). If the model is fixed-scale, use the corresponding value from the folder name.
• If your binary expects a model name (-n) instead of a folder, omit -n when using -m with a directory, or set -n to the base model name if required by your build.

Examples with bundled folders:

# Real-ESRGAN general x4
realesrgan-ncnn-vulkan -i in.png -o out.png -s 4 -m ./realesrgan-x4plus

# Real-ESRGAN anime x4
realesrgan-ncnn-vulkan -i in.png -o out.png -s 4 -m ./realesrgan-x4plus-anime

# RealESR AnimeVideo v3 (x3)
realesrgan-ncnn-vulkan -i in.png -o out.png -s 3 -m ./realesr-animevideov3-x3

# NMKD models (typically x4 or x8, match -s accordingly)
realesrgan-ncnn-vulkan -i in.png -o out.png -s 4 -m ./4x_NMKD-Superscale-SP_178000_G

Batch a whole folder:

realesrgan-ncnn-vulkan -i ./inputs -o ./outputs -s 4 -m ./realesrgan-x4plus

Performance tips:

• Use -t <threads> to control CPU threads used for pre/post steps
• Use -g <tiles> to adjust tiling for memory; smaller tiles reduce VRAM usage
• Add -f jpg or -f png to force output format

Directory layout suggestion

You can keep all models at the repo root (as provided) or move them under ./models/ and reference with -m ./models/<model-folder>.

Convert your own Models!

The Upscayl wiki has a guide on converting ESRGAN .pth models to NCNN (.param/.bin) for use as custom models:

• Model Conversion Guide: https://github.com/upscayl/upscayl/wiki/Model-Conversion-Guide

Credits and licenses

This repository aggregates model files from their respective authors. Please refer to the original projects for licenses and usage terms:

• Real-ESRGAN (Xintao et al.) — see the Real-ESRGAN project for details
• NCNN Vulkan ports and tools by nihui and contributors
• NMKD models — credits to NMKD for training and releases

If a specific subfolder includes its own LICENSE or README, that document takes precedence for that model. If you are an author and want adjustments or takedowns for your model files, please open an issue or PR.

Disclaimer

Models are provided as-is with no guarantees. Ensure you comply with each model’s license and any usage restrictions.