Add Swin2SR image upscaling Learning Path with Arm VGF

[usamahz]

Summary

Add an introductory Learning Path for upscaling a low-resolution image with Swin2SR, ExecuTorch, and Arm VGF on a Linux host. The walkthrough follows one end-to-end task: prepare the environment and image, export the pinned pretrained model, build and run the VGF host runner, and inspect the 64 × 64 to 128 × 128 result.

Review the introduction and four numbered pages in order, then the editable Excalidraw diagram and image comparison. Equal display sizes make detail differences visible, while labels distinguish actual pixel dimensions and the high-resolution reference from model input. The walkthrough uses FP32 export without calibration or smoke-test machinery.

The content retains draft: true and its draft cascade under the repository's intake guidance, pending technical and editorial review.

Validation

Executed the documented workflow from a fresh public ExecuTorch checkout and Python virtual environment in Ubuntu 24.04 AArch64. Environment installation, Arm SDK setup, model export, runner build, demo inference, and a second-image run passed. Both outputs are 128 × 128 RGB images. The demo result differs from a fresh PyTorch reference by at most 1 per 8-bit color channel.

Testing used Mesa llvmpipe software Vulkan with shaderFloat64, not a physical Mali GPU. Successful runs used a 12 GiB container limit within a 16 GiB Docker VM; smaller-memory attempts were OOM-killed. These are functional checks for the tested Linux AArch64 configuration, not GPU performance, broad image-quality, or Linux x86_64 validation. Upstream pip check warnings remain for missing auxiliary-tool dependencies and a TOSA/FlatBuffers version mismatch, although the documented export and runtime flow passed with upstream pins.

The Hugo 0.130.0 extended build passes. All six Learning Path pages render with draft content enabled, and local link targets and fragments validate. The repository style checker reports zero issues using its basic passive-voice fallback, and the image audit reports zero faulty images. YAML metadata, shell-block syntax, Excalidraw JSON, SVG XML, and PNG integrity checks pass.

Contribution checklist

• I have reviewed Create a Learning Path.
• I have checked my contribution for confidential information.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution under the terms of the Creative Commons Attribution 4.0 International License.

This pull request was authored with assistance from OpenAI Codex.