Skip to main content
Real-ESRGAN provides a comprehensive collection of pre-trained models optimized for different types of content. Choose the right model based on your input type and requirements.

Model Categories

General Models

Models optimized for general photographic images

Anime Models

Specialized models for anime images and illustrations

Video Models

Lightweight models designed for video upscaling

Quick Reference

General Images

ModelScaleUse Case
RealESRGAN_x4plus4xBest quality for general images
RealESRGAN_x2plus2xGeneral images with 2x upscaling
realesr-general-x4v31x-4xLightweight, fast processing
RealESRNet_x4plus4xMSE loss (smoother results)

Anime Content

ModelScaleUse Case
RealESRGAN_x4plus_anime_6B4xAnime images and illustrations
realesr-animevideov31x-4xAnime videos (lightweight)

Model Selection Guide

Use RealESRGAN_x4plus for the best quality on natural images, photographs, and general content. This is the flagship model with 23 RRDB blocks.If you need faster processing or have limited GPU memory, try realesr-general-x4v3, which is a compact model that still delivers good results.
Use RealESRGAN_x4plus_anime_6B for anime images and illustrations. This model is optimized specifically for anime characteristics and uses only 6 RRDB blocks for efficiency.For anime videos, use realesr-animevideov3, which is designed for temporal consistency and fast processing.
Video processing requires models that balance quality with speed. Use:
  • realesr-animevideov3 for anime videos
  • realesr-general-x4v3 for general video content
Both models support variable scaling (1x, 2x, 3x, 4x) and have compact architectures for efficient processing.
If you prefer smoother, less sharp results (useful for certain artistic styles), use RealESRNet_x4plus. This model is trained with MSE loss, producing over-smooth effects compared to GAN-based models.
Discriminator Models: Models ending with _netD are discriminator networks used for fine-tuning. You typically don’t need these for inference.

Download Locations

All models are hosted on GitHub Releases:
https://github.com/xinntao/Real-ESRGAN/releases
Models are automatically downloaded when using the inference scripts with the -n parameter:
python inference_realesrgan.py -n RealESRGAN_x4plus -i inputs

Next Steps

General Models

Detailed specifications and usage

Anime Models

Anime-specific optimizations

Video Models

Video processing workflows

Build docs developers (and LLMs) love