Anime Images

Upscale anime illustrations and artwork with the optimized RealESRGAN_x4plus_anime_6B model

Overview

Real-ESRGAN provides a specialized model optimized for anime images, illustrations, and artwork. The RealESRGAN_x4plus_anime_6B model is significantly smaller than the general model while delivering excellent results for anime content.

RealESRGAN_x4plus_anime_6B Model

Key Features

• Optimized for Anime: Trained specifically on anime and illustration data
• Smaller Size: Only 6 RRDB blocks vs. 23 blocks in the general model (~17MB vs ~64MB)
• Fast Processing: Faster inference due to smaller architecture
• High Quality: Preserves anime-style details and aesthetics

Download and Install

1. Download Model

   wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth -P weights
   
2. Run Inference

   python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i inputs
   
3. Check Results

   Enhanced anime images are saved in the results folder with 4x upscaling.

Usage Examples

Basic Upscaling

Single Image

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime_image.jpg

Advanced Options

With Tiling (Large Images)

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i large_anime.png --tile 400

Model Specifications

Property	Value
Architecture	RRDBNet with 6 blocks
Upscale Factor	4x
Model Size	~17MB
Input Channels	3 (RGB)
Output Channels	3 (RGB)
Feature Channels	64
Growth Channels	32
Training Data	Anime images and illustrations

Comparison with Other Models

vs RealESRGAN_x4plus

Feature	Anime 6B	x4plus General
Anime Quality	Excellent	Good
General Images	Good	Excellent
Model Size	17MB	64MB
Speed	Faster	Slower
Best For	Anime/Illustrations	Photos/Natural images

vs waifu2x

Real-ESRGAN anime model provides comparable or better quality than waifu2x with these advantages:

• More accurate detail reconstruction
• Better handling of compression artifacts
• Sharper edges and cleaner lines
• Native PyTorch implementation for GPU acceleration

See detailed comparisons in the source documentation.

vs realesr-animevideov3

Feature	Anime 6B	AnimeVideo v3
Image Quality	Excellent	Very Good
Video Quality	N/A	Excellent
Model Size	17MB	~8MB (XS)
Best For	Static images	Video frames

Use RealESRGAN_x4plus_anime_6B for static anime images and realesr-animevideov3 for video processing.

Common Use Cases

Upscaling Anime Screenshots

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i screenshots -o upscaled

Perfect for enhancing low-resolution anime screenshots from streaming services.

Processing Manga Panels

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i manga_pages -o enhanced --ext png

Preserves fine line details in manga artwork.

Enhancing Anime Artwork

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i artwork.jpg --outscale 4 --ext png

Upscale digital artwork and illustrations while maintaining anime aesthetic.

Processing Game Assets

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i game_assets -o enhanced_assets --tile 400

Upscale anime-style game sprites and textures. Use tiling for large textures.

Tips for Best Results

Image Format

Use PNG format for anime images when possible to avoid JPEG compression artifacts:

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i input.jpg --ext png

Transparent Images

The anime model handles RGBA images with transparency:

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime_with_alpha.png

Output is automatically saved as PNG to preserve the alpha channel.

Face Enhancement Not Recommended

The --face_enhance option is designed for realistic faces and may not work well with anime faces. It’s recommended to not use face enhancement with anime models:

# Don't use --face_enhance with anime images
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg

Choosing Output Scale

For anime images:

• 2x: Moderate enhancement, good for already decent quality
• 4x: Standard upscaling for most anime content
• Higher: Use for very low resolution sources

# For moderate quality anime
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --outscale 2

# For low quality anime
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --outscale 4

Performance Optimization

Memory Usage

The anime model uses less memory than the general model:

Standard Processing

# No tiling needed for most images
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg

Batch Processing

Process multiple anime images efficiently:

# Process entire folder
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime_collection -o upscaled_collection

# With custom settings
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime_collection \
  -o upscaled_collection --ext png --suffix enhanced

GPU Selection

Use specific GPU for processing:

# Use GPU 0
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --gpu-id 0

# Use GPU 1
python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --gpu-id 1

NCNN Executable Alternative

For users who prefer the portable NCNN executable:

1. Download NCNN Executable

   Download the appropriate version for your platform:

   • Windows
   • Linux
   • MacOS
2. Run NCNN Version

   ./realesrgan-ncnn-vulkan.exe -i anime.jpg -o output.png -n realesrgan-x4plus-anime
   

The NCNN version doesn’t require CUDA or PyTorch installation.

Troubleshooting

Results look over-smoothed

This is normal for the anime model as it’s trained to reduce noise. If you want to preserve more texture:

1. Use the general model instead:

   python inference_realesrgan.py -n RealESRGAN_x4plus -i anime.jpg
   

2. Or use a lower output scale:

   python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --outscale 2
   

Colors look different

Ensure you’re using PNG output for best color preservation:

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.jpg --ext png

CUDA out of memory

Use tiling to reduce memory usage:

python inference_realesrgan.py -n RealESRGAN_x4plus_anime_6B -i anime.png --tile 400

Try smaller tile sizes (256, 300) if still encountering issues.

Next Steps

Anime Videos

Learn to upscale anime videos with realesr-animevideov3

General Images

Explore models for real-world photos

NCNN Executable

Use the portable executable without Python