RealESR_Gx4_fp16

RealESR_Gx4_fp16

RealESR_Gx4_fp16 is an ONNX export of a Real-ESRGAN-based super-resolution model designed for general-purpose 4× image upscaling.

Model Description

This repository provides a float16 ONNX version of a Real-ESRGAN model optimized for a wide range of content, including natural images, video frames, and mixed visual styles.

This variant is intended as a balanced default for general upscaling tasks.

File

• RealESR_Gx4_fp16.onnx

Intended Use

This model is intended for:

• 4× image upscaling
• video frame enhancement
• general-purpose super-resolution workflows
• ONNX Runtime based applications

Input / Output

• Input: low-resolution image tensor
• Output: 4× upscaled image

Exact preprocessing and normalization depend on the host application.

Usage Notes

• Balanced for general content (photographic + mixed scenes)
• FP16 export for improved performance on supported hardware
• Compatible with ONNX Runtime (CUDA, TensorRT, DirectML, CPU)
• Suitable as a default upscaler in most pipelines

Source / Attribution

This repository contains an ONNX-packaged version of a Real-ESRGAN-based model.

The upstream Real-ESRGAN project is licensed under the BSD 3-Clause License.

Limitations

• GAN-based models may introduce artificial or hallucinated detail
• Not optimized specifically for anime or line-art (use anime variant instead)
• Output quality depends on input resolution and compression

License

BSD 3-Clause