General Images

Upscale real-world photos and general images with Real-ESRGAN

Overview

Real-ESRGAN provides several models optimized for general real-world images, including photos, screenshots, and natural scenes. This guide covers model selection and best practices for general image super-resolution.

Recommended Models

RealESRGAN_x4plus (Primary)

The default and most versatile model for general images.

Download Model

wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P weights

Specifications:

• Upscale Factor: 4x
• Architecture: RRDBNet (23 blocks)
• Model Size: ~64MB
• Best For: Photos, natural images, general scenes

realesr-general-x4v3 (Denoise Control)

A smaller, faster model with controllable denoising strength.

Download Models

wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth -P weights
wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-wdn-x4v3.pth -P weights

Specifications:

• Upscale Factor: 4x
• Architecture: SRVGGNetCompact (32 conv layers)
• Model Size: ~17MB (much smaller than x4plus)
• Best For: Images where you want to control noise vs. detail trade-off
• Denoise Strength: 0 (keep noise/texture) to 1 (strong denoise)

The realesr-general-x4v3 model requires both the base model and the denoise model (wdn) for the -dn option to work.

RealESRGAN_x2plus (2x Upscaling)

For moderate upscaling when 4x is too aggressive.

Download Model

wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth -P weights

Specifications:

• Upscale Factor: 2x
• Architecture: RRDBNet (23 blocks)
• Best For: Images that need moderate enhancement

RealESRNet_x4plus (No GAN)

ESRGAN model without GAN training - produces smoother results.

Download Model

wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.1/RealESRNet_x4plus.pth -P weights

Specifications:

• Upscale Factor: 4x
• Architecture: RRDBNet (23 blocks)
• Best For: When you want smoother results without GAN artifacts

Model Comparison

By Use Case

Use Case	Recommended Model	Notes
General photos	RealESRGAN_x4plus	Best overall quality
Photos with faces	RealESRGAN_x4plus + --face_enhance	Combines super-resolution with GFPGAN
Noisy images	realesr-general-x4v3 with high --denoise_strength	Control noise reduction
Textured images	realesr-general-x4v3 with low --denoise_strength	Preserve texture details
Quick processing	realesr-general-x4v3	Smaller, faster model
Moderate upscaling	RealESRGAN_x2plus	2x instead of 4x
Smooth results	RealESRNet_x4plus	No GAN artifacts

By Size

Model	Size	Speed	Quality
realesr-general-x4v3	~17MB	Fast	Good
RealESRGAN_x2plus	~64MB	Medium	Very Good
RealESRGAN_x4plus	~64MB	Medium	Excellent
RealESRNet_x4plus	~64MB	Medium	Excellent (smooth)

Common Usage Patterns

Standard Workflow

1. Prepare Images

   Place your images in the inputs folder:

   mkdir -p inputs
   cp /path/to/your/images/* inputs/
   
2. Run Inference

   Process with the appropriate model:

   python inference_realesrgan.py -n RealESRGAN_x4plus -i inputs -o results
   
3. Check Results

   Enhanced images are saved in the results folder with the _out suffix.

Advanced Examples

Custom Scale

# Upscale to exactly 3.5x
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --outscale 3.5

Tips for Best Results

CUDA Out of Memory

For high-resolution images, use the --tile option:

python inference_realesrgan.py -n RealESRGAN_x4plus -i large_image.jpg --tile 400

Smaller tile sizes use less memory but may introduce slight artifacts at tile boundaries.

Denoise Strength Guidelines

When using realesr-general-x4v3:

• 0.0-0.3: Preserve original texture and noise (good for film grain)
• 0.4-0.6: Balanced (default is 0.5)
• 0.7-1.0: Strong noise reduction (good for compressed images)

Arbitrary Output Scale

Use --outscale for any target size:

# Exactly 1.5x
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --outscale 1.5

# Exactly 10x
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --outscale 10

The model upscales to its native scale (e.g., 4x), then resizes to your target using LANCZOS4.

Batch Processing

Process entire folders efficiently:

# Process all images in a folder
python inference_realesrgan.py -n RealESRGAN_x4plus -i input_folder -o output_folder

# With custom settings
python inference_realesrgan.py -n RealESRGAN_x4plus -i input_folder -o output_folder \
  --outscale 4 --tile 400 --ext png

The script automatically processes all supported images in the input folder.

Special Image Types

RGBA Images (With Transparency)

Real-ESRGAN automatically handles alpha channels:

python inference_realesrgan.py -n RealESRGAN_x4plus -i image_with_alpha.png

• Alpha channels are upsampled using the method specified by --alpha_upsampler (default: realesrgan)
• Output is always saved as PNG to preserve transparency

Grayscale Images

Grayscale images are supported natively:

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

16-bit Images

High bit-depth images are automatically handled:

python inference_realesrgan.py -n RealESRGAN_x4plus -i 16bit_image.png

Performance Optimization

Multi-GPU Usage

Specify GPU device:

# Use GPU 0
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --gpu-id 0

# Use GPU 1
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --gpu-id 1

Memory Management

Low Memory

# Use smaller tiles and general model
python inference_realesrgan.py -n realesr-general-x4v3 -i input.jpg --tile 256

Balanced

# Default settings with moderate tiling
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --tile 400

High Memory

# No tiling for best quality (requires sufficient VRAM)
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg

Next Steps

Face Enhancement

Enhance faces in your photos with GFPGAN integration

Anime Images

Learn about specialized models for anime content