Skip to content

super-resolution

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

super-resolution

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
assets
assets
 
 
README.md
README.md
 
 

README.md

Ryzen™ AI Super-Resolution

Running Super-Resolution Models on Ryzen AI

This Ryzen AI example lets you upscale images using state-of-the-art super-resolution models running locally on your AMD NPU. Two model families are provided:

The models have been quantized to INT8 and optimized for the AMD Ryzen AI NPU using ONNX Runtime with the Vitis AI Execution Provider.

The goal of this example is to show you how to run inference of these super-resolution models on an AMD NPU.

Prerequisites

Hardware

You need an AMD AI PC with a Ryzen AI NPU. Supported processors:

Series Codename Abbreviation Launch Year Windows 11
Ryzen AI Max PRO 300 Series Strix Halo STX 2025 ✔️
Ryzen AI PRO 300 Series Strix Point / Krackan Point STX/KRK 2025 ✔️
Ryzen AI Max 300 Series Strix Halo STX 2025 ✔️
Ryzen AI 300 Series Strix Point STX 2025 ✔️

Software

  1. Install the latest NPU drivers and Ryzen AI Software by following the Ryzen AI SW Installation Instructions. Allow approximately 30 minutes for the full installation.

  2. Clone and activate your own conda environment based on the installed ryzen-ai-v.v.v environment installed. Substitute the correct version number for v.v.v (e.g., 1.7.1). Also, be sure to set the environment variable for the RYZEN_AI_INSTALLATION_PATH.

conda create -n super-resolution --clone ryzen-ai-1.7.1
conda activate super-resolution
$Env:RYZEN_AI_INSTALLATION_PATH = 'C:/Program Files/RyzenAI/1.7.1/'

Real-ESRGAN — 4x Super-Resolution

Real-ESRGAN (Real Enhanced Super-Resolution Generative Adversarial Networks) produces high-quality 4x upscaled images with strong perceptual fidelity. This AMD version has been re-trained with reduced feature channels and fewer stacked blocks for improved efficiency, then quantized to INT8 for the AMD NPU. Here is an example input and output image produced with the 256x256 tile variant of the model.

Input image Output image
assets/input_tiger_320x480_108005.png assets/output_tiger_4x_1280x1920_108005.png

Figure 1: Input 320x480 scaled up by 4x to 1280x1920 with Real-ESRGAN model running on AMD AI PC NPU. Image source: EDSR Benchmark dataset (edsr_benchmark\B100\HR\108005.png).

For more details on the tiling method, please visit the AMD blog: Super Resolution Acceleration on Ryzen AI NPU.

Four tile-size variants are available. All share the same code and inference pipeline — only the ONNX model file differs:

Model Tile Size Hugging Face Link
Real-ESRGAN 128×128 amd/realesrgan-128x128-tiles-amdnpu
Real-ESRGAN 256×256 amd/realesrgan-256x256-tiles-amdnpu
Real-ESRGAN 512×512 amd/realesrgan-512x512-tiles-amdnpu
Real-ESRGAN 1024×1024 amd/realesrgan-1024x1024-tiles-amdnpu

Tile size tradeoffs: A larger tile size reduces stitching overhead and may produce fewer boundary artifacts, but requires more memory per tile. A smaller tile size is more memory-friendly and can be better suited for lower-resolution inputs.

Quick Start

The example below uses the 256×256 variant. To use a different tile size, substitute the repository name and ONNX filename accordingly.

  1. Clone the model repository from Hugging Face:
git clone https://hf.co/amd/realesrgan-256x256-tiles-amdnpu
cd realesrgan-256x256-tiles-amdnpu
  1. Install dependencies (you should already be in the activated super-resolution conda environment):
pip install -r requirements.txt
  1. Run inference on an image:
python onnx_inference.py --onnx onnx-models/realesrgan_nchw_256x256_u8s8.onnx --input assets/input_tiger_320x480_108005.png --out-dir outputs --device npu

For other tile sizes, visit the other Hugging Face repos and follow the same steps.

Arguments

Argument Description
--onnx Path to the ONNX model file
--input A single image file or a directory (recursively processes .png, .jpg, .jpeg)
--out-dir Output directory for the upscaled images
--device npu (default) or cpu

To use evaluation scripts, benchmark datasets, and accuracy metrics, see the full model card on Hugging Face: amd/realesrgan-256x256-tiles-amdnpu.

SESR-M7 — 2x Super-Resolution

SESR (Super-Efficient Super Resolution) is a lightweight CNN-based model designed for fast, efficient image upscaling. The SESR-M7 is one of the smaller variants of SESR and uses a tiling technique to work across the image in 256x256 pixel tiles to produce a 2x super-resolution output. Even though the tile size is 256×256, the inference pipeline handles images of almost any size. For example, here is an input of 339x510 and the 2x output of 678x1020.

For more details on the tiling method, please visit the AMD blog: Super Resolution Acceleration on Ryzen AI NPU.

Input image Output image
assets/input_ice_climber_0844.png assets/output_ice_climber_0844_x2.png

Figure 2: Ice climber image upscaled by 2x with SESR-M7 model running on AMD AI PC NPU. Source: DIV2K dataset (DIV2K_valid_LR_bicubic\X4\0844x4.png).

Quick Start

  1. Clone the model repository from Hugging Face:
git clone https://hf.co/amd/sesr-m7-256x256-tiles-amdnpu
cd sesr-m7-256x256-tiles-amdnpu
  1. Install dependencies (you should already be in the activated super-resolution conda environment):
pip install -r requirements.txt
  1. Run inference on an image:
python onnx_inference.py --onnx onnx-models/sesr_nchw_int8.onnx --input assets/input_ice_climber_0844.png --out-dir outputs --device npu

Arguments

Argument Description
--onnx Path to the ONNX model file
--input A single image file or a directory (recursively processes .png, .jpg, .jpeg)
--out-dir Output directory for the upscaled images
--device npu (default) or cpu

To use evaluation scripts, benchmark results, and training details, see the full model card on Hugging Face: amd/sesr-m7-256x256-tiles-amdnpu.

Notes

  • First-run compilation: If the model has not been previously compiled for your NPU, the first inference run will take several minutes to compile and cache the model. Subsequent runs will be fast. The compiled model cache is stored in the modelcachekey_* directory within the repository.
  • CPU fallback: All models can also run on CPU by passing --device cpu. No NPU drivers or Ryzen AI SW are required for CPU execution, but performance will be significantly slower.

Additional Resources

Join the AMD AI Developer Program to get access to tools, resources, and a community of developers building on AMD hardware.