waifu2x Review: The 28k-Star Anime Super-Resolution Tool That Saves Low-Res Art

I’m a long-time anime fan with a hard drive full of screenshots and wallpapers collected over a decade. The problem? Many old images are low-resolution and look terrible on a 4K monitor. I’ve tried plenty of upscaling tools, and waifu2x turned out to be the best solution for anime art.

Project Background

Developed by nagadomi and released in 2015, waifu2x is one of the early landmark projects that applied deep learning to image super-resolution. It has 28k stars on GitHub. While the last update was in May 2023, the core model remains effective today.

It’s built on the Torch/Lua framework and specifically optimized for anime-style artwork. License is MIT, so you’re free to use and modify it.

Where It Genuinely Excels

1. Unmatched Line Art Restoration

waifu2x’s core strength isn’t just pixel upscaling — it’s specifically tuned for the clean lines and flat color regions characteristic of anime art. Traditional upscaling methods (like bicubic interpolation) blur lines and create jagged edges, but waifu2x restores sharpness and cleanliness.

I tested a 720p screenshot from a 2006 series upscaled to 1920x1080. Character outlines stayed crisp, color boundaries didn’t bleed, and the effect held up on landscape wallpapers too — sky gradients transitioned smoothly without banding.

2. Denoising and Upscaling Combined

Many old images suffer from both low resolution and compression artifacts. waifu2x handles denoising and upscaling simultaneously, saving you from preprocessing with separate tools.

It offers three denoising levels: low, middle, and high. My general strategy:

• Clean sources (official artwork): no denoising, 2x upscale
• Standard screenshots: low denoising + 2x upscale
• Heavily compressed old images: medium denoising + 2x upscale

I rarely use the high setting because aggressive denoising sacrifices detail, giving images an oil-painting look.

3. Local Processing, Privacy Guaranteed

All processing happens locally — your images never leave your machine. For someone like me with a massive collection of personal screenshots, this is non-negotiable. Many online upscaling tools require uploading to cloud servers, which always feels risky.

4. Multiple Ways to Use It

waifu2x offers several interfaces to match different needs:

Online demo: Visit https://waifu2x.udp.jp/, drag and drop your image, and process it right in the browser. Great for occasional use.

Local command line:

# Requires torch and nn dependencies
th waifu2x.lua -m noise_scale -noise_level 1 -i input.png -o output.png

Third-party GUI tools: The community has created many user-friendly wrappers, such as waifu2x-caffe (Windows) and waifu2x-ncnn-vulkan (cross-platform). I personally use waifu2x-ncnn-vulkan because it supports GPU acceleration, making processing much faster.

5. Flexible Scale Factors

Supports 1x (denoise only) and 2x upscaling. For larger sizes, you can chain multiple 2x passes (e.g., 2x followed by another 2x for 4x total). Just note that quality degrades slightly with each pass, so I wouldn’t go beyond 4x in practice.

But It Has Real Limitations

Anime-only territory. waifu2x was trained exclusively on anime images, and it shows. Throw a real-world photograph at it, and you’ll get weird smearing and unnatural detail recovery. Tools like Real-ESRGAN or Topaz Gigapixel AI are far better for photos. If you primarily work with photographs, skip this one.

Setup can be tricky. The official release requires a Lua + Torch environment, which isn’t beginner-friendly. Fortunately, the community has packaged plenty of GUI versions, so casual users can just grab waifu2x-ncnn-vulkan or use the online version.

Slow on pure CPU. Using the official Lua version on CPU, upscaling a 1920x1080 image to 2x can take several minutes. Switch to a GPU-accelerated or ncnn-optimized build for 10x+ speedups. My RTX 3060 handles the same image in 3-5 seconds with waifu2x-ncnn-vulkan.

Limited color mode support. Certain PNG formats (like indexed-color PNGs with alpha) may throw errors and need conversion to true-color RGB first. I’ve run into this a few times — a quick conversion in Photoshop or ImageMagick fixes it.

Maintenance has slowed. The core author’s last update was in 2023. While the model itself doesn’t need frequent updates, platform compatibility improvements now rely mostly on community efforts.

How It Compares

Real-ESRGAN: Its anime model (animevideo) comes close to waifu2x, and it handles real-world photos much better. If you process both types of images, Real-ESRGAN is more versatile. But for pure anime content, waifu2x still produces cleaner line art.

Topaz Gigapixel AI: A commercial tool with the best overall quality, supporting both photos and anime. But it costs $199 and requires an internet connection for license validation. waifu2x is free and open-source, making it the obvious choice for budget-conscious users.

Original ESRAN: Inferior to waifu2x for anime, tending to over-sharpen and produce artifacts. Real-ESRGAN is the improved successor.

waifu2x-caffe: A Windows-only GPU-accelerated build based on the Caffe framework. If you have an NVIDIA card and run Windows, this is among the fastest local processing options available.

Who It’s For

Long-time anime fans with collections of low-res screenshots, wallpaper collectors, AMV/MAD creators who need to restore source materials, and anyone who needs to upscale anime-style artwork.

My current workflow: old images first go through waifu2x for 2x upscale + low denoising. If I need even higher resolution, I run them through Real-ESRGAN as a second pass. Together they cover virtually all my anime image restoration needs.

Bottom Line

waifu2x is a classic in the anime super-resolution space. Nearly a decade after release, its core model still holds its own for pure anime content. Those 28k stars represent the appreciation of countless fans who’ve used it to bring blurry childhood memories back into focus.

If your hard drive is full of 480p and 720p old-series screenshots, spend ten minutes downloading waifu2x-ncnn-vulkan. You’ll thank me later.

GitHub: https://github.com/nagadomi/waifu2x Online demo: https://waifu2x.udp.jp/

---

About the Author

Liudingyu is a full-stack developer and heavy GitHub user. With 900+ starred repos over the past 3 years, this site only covers tools I’ve actually used or deeply researched.

📧 Found a great tool to recommend? Email [email protected]