Skip to the content
Know your tech before you buy it. Makers: get listed from $49/mo

Real-CUGAN

github.com

  • Free plan
  • Windows
  • macOS
  • Linux
  • Web
  • Android
  • Self-hosted
8.2 Club score Visit github.com

What Real-CUGAN is

Real-CUGAN is a free AI super-resolution model made for anime images. It supports 2×, 3×, and 4× enlargement, with denoising strengths, a conservative mode, and adjustable alpha. Its 1× model is still in training. Different enhancement strengths are available through five model weights for 2× and three each for 3× and 4×. The Windows setup handles images and video, and the repository includes a video inference script. Windows packages are linked through Baidu Drive, GitHub Releases, and Google Drive; the package was tested on 64-bit Windows 10. Nvidia use requires 1.5 GB of video memory, with different CUDA minimums for light and heavy versions. A portable NCNN option has Windows, Linux, and macOS executables for Intel, AMD, Nvidia, and Apple-Silicon GPUs using Vulkan; it includes binaries and models and does not need CUDA or PyTorch runtime. Browser options include a demo and Web-CPU version, though the CPU demo is slow and smaller images are recommended. Cache and tile settings offer ways to reduce GPU-memory use.

Who it suits

Real-CUGAN suits people enlarging anime images who want several enhancement strengths or image and video processing. Its NCNN option may suit users with Intel, AMD, Nvidia, or Apple-Silicon GPUs who want to avoid CUDA and PyTorch runtimes.

Good to know, and what to watch

  • Free model and downloadable packages.
  • Supports 2×, 3×, and 4× upscaling.
  • Windows configuration handles images and video.
  • Portable NCNN package needs no CUDA or PyTorch runtime.
  • Cache and tile settings can reduce GPU-memory use.
  • The 1× model is still in training.
  • The CPU demo is slow.
  • Training code is not public.

Real-CUGAN

Free

Free plan

The verdict

Real-CUGAN offers several anime upscaling strengths and routes for Windows, other desktop platforms, and browser use. Check the applicable package requirements and GPU-memory settings before choosing a workflow.

What the maker publishes

Alternative GPU support
The linked NCNN implementation offers Windows, Linux, and macOS executables for Intel, AMD, Nvidia, and Apple-Silicon GPUs using Vulkan. source
Desktop app
Yes source
Download
The maker links Windows executable packages through Baidu Drive, GitHub Releases, and Google Drive. source
Enhancement models
The English README says 2× has five model weights and 3×/4× have three model weights each for different enhancement strengths. source
Free plan
Yes source
Image and video
The Windows configuration supports both image and video processing, and the repository includes a simple video inference script. source
Integration
The project describes its model architecture as compatible with Waifu2x-CUNet, and it lists PyTorch and VapourSynth use paths. source
Memory limits
The Windows package provides cache and tile settings to trade processing speed for lower GPU-memory use, including modes intended for very large images. source
Portable NCNN package
The NCNN package includes binaries and models and states that it needs no CUDA or PyTorch runtime. source
Purpose
Real-CUGAN is an AI super-resolution model trained on a million-scale anime dataset and designed for anime images. source
Security and compliance
The maker README and repository page provide no security or compliance claims. source
Support
The maker invites users to leave suggestions and requests in the repository issues. source
Training code
The maker README says the training code is not public. source
Upscaling
It supports 2×, 3×, and 4× upscaling; the 1× model is described as still in training. source
Web option
The maker README links a browser demo and a Web-CPU version; the README notes that the CPU demo is slow enough that smaller images are recommended. source
Windows requirements
The package was tested on Windows 10 64-bit; Nvidia use requires 1.5 GB of video memory, with different CUDA minimums for light and heavy versions. source

Checked against the maker's pages on 9 Oct 2026.