Forum Discussion
Snapdragon AI Studio – Native Local AI Image Generation for Windows 11 ARM64
If you have experience with C++, Python, ONNX Runtime, or Qualcomm's QNN SDK, your pull requests are welcome. The developer has identified key focus areas for RC2, including image quality and performance.
Thanks, Sallop. I appreciate you pointing this out. Contributions are definitely welcome, especially around performance, image quality and the QNN/NPU side of the project. RC2 has moved quite a bit forward since the original post, but there is still plenty to explore and improve.
Holger Kreuzhofen
Founder & Lead Developer
Snapdragon AI Studio – Phoenix Engine
RC2:
https://github.com/Kreuzhofen/snapdragon-ai-studio/releases/tag/v2.0.0-rc.2
RC2A update: Snapdragon AI Studio – Phoenix Engine 2.0 RC2A is now available.
Since my RC2 update, I continued working on model setup and reliability, with a particular focus on Stable Diffusion 3.5 Medium.
I'm happy to report that the Stable Diffusion 3.5 Medium workflow using Qualcomm QAI AppBuilder is now working end-to-end in the real user flow on my Snapdragon system:
Not installed → guided setup → Qualcomm download → preparation → import → validation → activation → real image generation.
RC2A also further improves model download and installation, including handling of incomplete or outdated setup data, validation and automatic model activation.
🔥 Phoenix Boost has also received another round of improvements. The guided setup, local Ollama/Qwen2.5 3B integration, live progress reporting and inference reliability have all been refined.
The goal remains the same: make local AI image generation on Windows on Snapdragon easier to install and use without requiring users to deal with the technical details behind the models and runtimes.
Updated user documentation is now also available in English, German and Spanish.
RC2A:
https://github.com/Kreuzhofen/snapdragon-ai-studio/releases/tag/v2.0.0-rc.2a
Feedback and testing are very welcome.
Holger Kreuzhofen
Founder & Lead Developer
Snapdragon AI Studio – Phoenix Engine
Snapdragon AI Studio – Phoenix Engine is an independent open-source project and is not an official Qualcomm product.
Further update: Snapdragon AI Studio – Phoenix Engine 2.0 RC2B is now available.
RC2B is the most stable, refined, and user-friendly release candidate to date. The focus of this update is a more reliable and polished experience for running local generative AI on Windows 11 ARM64 devices powered by Snapdragon processors.
Key improvements include:
• A more responsive and refined Phoenix interface
• Improved first-run experience and model management
• More reliable local model installation workflows
• Validated Stable Diffusion image-generation pipelines
• Qualcomm QNN and Snapdragon NPU acceleration
• Optional local prompt refinement through Phoenix Boost
• Additional stability, diagnostics, and usability improvements
The main objective remains unchanged: make local AI image generation on Windows on Snapdragon easier to install, understand, and use without requiring users to manage the underlying runtime complexity manually.
RC2B is available as an open-source pre-release for testing and evaluation:
https://github.com/Kreuzhofen/snapdragon-ai-studio/releases/tag/v2.0.0-rc.2b
Constructive feedback and reproducible technical issue reports are very welcome.
Kind regards,
Holger Kreuzhofen
Founder & Lead Developer
Snapdragon AI Studio – Phoenix Engine
Quick update: HK NPU STUDIO 2.0 RC3 is now available.
The project has evolved significantly since my original post, so I also decided to give it a clearer identity:
Snapdragon AI Studio is now HK NPU STUDIO.
The technical direction remains the same: local AI workloads for Windows 11 ARM64 and Snapdragon® PCs, with a strong focus on Qualcomm QNN/HTP and NPU acceleration.
RC3 introduces AI Photo Restoration to Phoenix Image Lab.
The new restoration workflow is designed for old, damaged, low-resolution, and historical photographs, with a strong focus on preserving the original image as faithfully as possible.
Highlights in RC3:
- New AI Photo Restoration workflow
- Faithful Restore for source-preserving restoration
- Up to 4× upscaling
- Improved preservation of faces, hair, clothing, textures, and fine details
- Dust Cleanup V4 for conservative removal of dust, specks, and small defects
- Native multi-tile processing for larger images
- Grayscale photographs remain grayscale
- Photo Restore models are included with the installer
- Neural restoration runs locally on the Snapdragon® NPU through Qualcomm QNN/HTP
- No CPU or GPU AI fallback is used for Photo Restoration
RC3 also includes improvements to memory handling, Windows 11 ARM64 startup robustness, QNN path handling, and installer packaging.
Download:
https://github.com/Kreuzhofen/hk-npu-studio/releases/tag/v2.0.0-rc.3Your Hardware • Your AI • Your Control
Holger Kreuzhofen
Founder & Lead Developer
HK NPU STUDIO