Real-ESRGAN

Real-ESRGAN (Real-World Enhanced Super-Resolution Generative Adversarial Network) is a blind image super-resolution model developed by Tencent's Applied Research Center (ARC) in collaboration with the Shenzhen Institutes of Advanced Technology (SIAT). Released in July 2021, the model is designed to restore and upscale low-resolution images by removing complex, unknown real-world degradations. It builds upon the classic ESRGAN framework but shifts the focus toward practical, general-purpose real-world applications using a novel data synthesis pipeline.

High-Order Degradation Modeling

The primary innovation of Real-ESRGAN is its high-order degradation modeling process. Standard super-resolution models typically train on simple, synthetic downsampled images, which fails to prepare them for complex real-world artifacts. Real-ESRGAN addresses this by introducing a multi-step degradation pipeline that applies various combinations of blur, noise, downsampling, and JPEG compression in successive stages. It also simulates digital artifacts such as ringing and overshoot. This rigorous synthetic degradation training allows the network to generalize well to real photographs and digital art without needing paired real-world training datasets.

Model Architecture

The architecture of Real-ESRGAN is composed of a generator and a discriminator:

• Generator: The generator network is based on the Residual-in-Residual Dense Block (RRDB) architecture inherited from ESRGAN. The standard RealESRGAN_x4plus model contains approximately 16.7 million parameters.
• Discriminator: Unlike the original ESRGAN which uses a VGG-style discriminator, Real-ESRGAN adopts a U-Net discriminator equipped with spectral normalization. The U-Net design allows the discriminator to evaluate local textures and fine details more effectively, which stabilizes the training dynamics of the generative adversarial network and suppresses unwanted artifacts.

Model Variants

To accommodate different kinds of content, Tencent released several specialized variants of the model:

• RealESRGAN_x4plus: The default model trained on a wide variety of real-world images, suitable for general photographs.
• RealESRGAN_x4plus_anime_6B: A streamlined 6-block variant designed specifically for anime, illustrations, and digital paintings, providing faster inference times and cleaner line-art upscaling.
• RealESRNet_x4plus: A non-adversarial variant trained solely with L1 loss, which produces fewer hallucinated details but offers faster, more stable outputs.
• realesr-animevideov3: A highly compact model optimized for super-resolution in animated video sequences.

For optimal results on portrait photographs, Real-ESRGAN is frequently coupled with face restoration models like GFPGAN or CodeFormer to reconstruct natural-looking facial details that general-purpose upscaling models might blur or distort.