Skip to main content
Find answers to common questions about using Real-ESRGAN, including model selection, face enhancement, and error resolution.

​
Model Selection

Different models are optimized for different types of content:
  • RealESRGAN_x4plus: General purpose model for real-world images
  • RealESRGAN_x4plus_anime_6B: Optimized for anime and illustration images
  • RealESRGANv2-animevideo-xsx2: For anime videos at 2x upscaling
  • RealESRGANv2-animevideo-xsx4: For anime videos at 4x upscaling
  • RealESRNet_x4plus: Suitable when you want less texture enhancement
For detailed model comparisons and recommendations, refer to the Model Zoo documentation.

​
Face Enhancement

The face_enhance option is only designed for real human faces and should not be used for anime content.
The face enhancement feature uses models trained specifically on real human faces and will not work properly on anime characters or illustrations. Using this option on anime content will:
  • Waste GPU memory unnecessarily
  • Not improve the output quality
  • Potentially introduce artifacts
Recommendation: Omit the --face_enhance flag when processing anime images or animation videos to save GPU memory and avoid potential issues.

​
Common Errors

This error occurs when running Real-ESRGAN inference on CPU with half precision (fp16), which is the default setting for memory efficiency.

​
Error Message

RuntimeError: "slow_conv2d_cpu" not implemented for 'Half'

​
Cause

Real-ESRGAN uses half precision (fp16) by default to:
  • Reduce GPU memory consumption
  • Speed up inference
However, some operators for half precision inference are not implemented in CPU mode.

​
Solution

Add the --fp32 option to your command to use full precision (fp32) instead:
python inference_realesrgan.py -n RealESRGAN_x4plus -i inputs --fp32
Or with the executable:
./realesrgan-ncnn-vulkan -i input.jpg -o output.png --fp32
Using --fp32 will increase memory usage but is required for CPU inference.
If you encounter GPU out-of-memory errors, try these solutions:
  1. Use tile mode: Process the image in smaller tiles 
    python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --tile 200
  2. Disable face enhancement (if enabled): 
    python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg
# Remove --face_enhance flag
  3. Reduce tile size: Lower the tile parameter value 
    python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg --tile 100
Smaller tile sizes use less memory but may take longer to process.

​
Installation Issues

After installing Real-ESRGAN, test your installation with a simple command:
python inference_realesrgan.py -n RealESRGAN_x4plus -i inputs
If you see output images generated without errors, your installation is working correctly.For dependency issues, ensure you have:
  • Python 3.7+
  • PyTorch 1.7+
  • basicsr
  • facexlib (if using face enhancement)
  • gfpgan (if using face enhancement)
Real-ESRGAN is compatible with Python 3.7 or higher.Check your Python version:
python --version
If you have multiple Python versions installed, you may need to use python3 explicitly:
python3 inference_realesrgan.py -n RealESRGAN_x4plus -i inputs

​
Performance

Several options can improve Real-ESRGAN’s inference speed:
  1. Use GPU acceleration: Ensure CUDA is properly installed
  2. Use the ncnn implementation: Faster for CPU inference 
    ./realesrgan-ncnn-vulkan -i input.jpg -o output.png
  3. Disable unnecessary features:
    • Remove --face_enhance if not needed
    • Use appropriate tile sizes (larger = faster but more memory)
  4. Choose the right model: Smaller models like RealESRGANv2 variants may be faster
The ncnn implementation offers better performance on both CPU and mobile devices.

Build docs developers (and LLMs) love

Get started for freeTalk to us