Releases: Kreuzhofen/hk-npu-studio
Release list
HK NPU STUDIO 2.0 RC2B
HK NPU STUDIO 2.0 RC2B
HK NPU STUDIO 2.0 RC2B is the current Windows 11 ARM64 release candidate for local AI workflows on Snapdragon® PCs.
🔄 Rebranding note
Snapdragon AI Studio has been renamed to HK NPU STUDIO. The project and development continue under the new name. Existing user data from previous installations is supported through the built-in migration path.
🌟 Highlights
- Native local AI image generation on Windows 11 ARM64
- Snapdragon NPU / QNN accelerated workflows
- Phoenix Boost prompt enhancement
- Responsive UI with high-DPI support
- Local model management
- Gallery and comparison workflows
- RealESRGAN NPU upscaling with 2× and 4× output
- Improved installer and application-data migration
- Legacy Snapdragon AI Studio user data migration support
📦 Windows ARM64 Installer
HKNPUStudio-2.0.0-rc.2b-ARM64-Setup.exe
SHA-256:
8CBAACD055B7CABFBEDBAF0756A31C8383AF5690A619F26D135DF795292EB459
🔎 Validation
Validated on the Windows ARM64 development system and RC2CleanTest.
Testing included clean installation, model installation and execution, image generation, Phoenix Boost, gallery workflows, RealESRGAN NPU upscaling, high-DPI layouts, and Light/Dark themes.
This remains a pre-release / release candidate intended for testing and evaluation.
Your Hardware • Your AI • Your Control
Local AI for Snapdragon® PCs
Featuring Phoenix Boost
Snapdragon is a trademark or registered trademark of Qualcomm Incorporated.
Snapdragon AI Studio 2.0 RC2A
Snapdragon AI Studio 2.0 RC2A
🔥 Phoenix Boost, a stronger installation experience, and Stable Diffusion 3.5 Medium
RC2A continues the major redesign introduced with RC2 and focuses on making Snapdragon AI Studio more reliable and easier to set up on a real Windows on Snapdragon system.
This release further improves Phoenix Boost, significantly hardens model download and installation, and completes an important milestone for Stable Diffusion 3.5 Medium using Qualcomm QAI AppBuilder.
🔥 Phoenix Boost — further refined
- Improved guided Phoenix Boost setup
- More reliable Ollama integration
- Local Qwen2.5 3B setup and prompt enhancement
- Improved real-time setup and download progress
- More robust Qwen inference handling and timeout behavior
- Prompt preview before generation
- Deterministic Boost remains independently available
- Phoenix Boost remains optional and does not block normal image generation
🧠 Stable Diffusion 3.5 Medium + Qualcomm QAI AppBuilder
RC2A substantially improves the Stable Diffusion 3.5 Medium installation workflow.
Phoenix now guides and automates the QAI AppBuilder workflow as far as possible:
- Locate the QAI AppBuilder ZIP
- Extract and prepare the Qualcomm workspace
- Prepare required dependencies
- Run the Qualcomm SD3.5 workflow
- Download the required model data
- Import the resulting model into Snapdragon AI Studio
- Create and verify installation metadata
- Validate the final model
- Activate it for generation
Existing extraction data is checked against the ZIP actually being used. Stale or mismatched workspaces can be replaced instead of silently being reused.
✅ Full real-world SD3.5 validation
The complete RC2A SD3.5 user flow has been successfully validated on real Snapdragon hardware:
Not installed → guided setup → Qualcomm model download → preparation → import → validation → activation → real image generation
The goal remains simple:
Select the model, let Phoenix prepare it, and create an image.
📥 More reliable download & installation
- More reliable download and startup behavior
- Better handling of existing installation data
- Detection of incomplete or damaged model sources
- Explicit redownload handling when required
- Stronger final model validation
- Automatic activation after successful installation
- Clearer preparation, download, validation and activation states
- Less technical decision-making for the user
🧭 Guided setup
RC2A continues the goal introduced with RC2:
Install → select model → guided setup → validate → activate → generate
Users should not need to understand individual QNN packages, ONNX components or Qualcomm model directory structures to get started.
📚 Documentation
Complete RC2A user guides are available in three languages:
⚡ Built for Windows on Snapdragon
- Windows 11 ARM64
- Snapdragon X Plus / X Elite
- Qualcomm Hexagon NPU
- Qualcomm QNN
- Qualcomm QAI AppBuilder
- ONNX Runtime
- Local AI image generation
- Privacy-first workflows
Snapdragon AI Studio is an independent open-source project and is not an official Qualcomm product.
Snapdragon AI Studio 2.0 RC2
Snapdragon AI Studio – Phoenix Engine 2.0 RC2
⚠️ Known Issue – Stable Diffusion 3.5 Medium
The guided first-time installation of Stable Diffusion 3.5 Medium may currently fail on a clean Windows 11 ARM64 system during local runtime preparation.
This issue is isolated to the SD3.5 installation workflow and does not affect Snapdragon AI Studio RC2 as a whole.
A fix is currently being prepared. Until it is available, users are advised not to start a new SD3.5 installation from the Model Manager.
Existing working model configurations are not affected by this notice.
Status: Fix in progress.
🚀 A Major Step Forward from RC1
Snapdragon AI Studio – Phoenix Engine 2.0 RC2 is a major evolution of the project and a significant step forward from RC1.
RC2 is not simply a collection of bug fixes. It substantially improves the complete experience — from first launch and model installation to local image generation on Snapdragon-powered Windows PCs.
The guiding goal behind RC2 has been simple:
Install Snapdragon AI Studio → select a model → generate an image.
Users should not need to understand model packages, tokenizer files, runtime components, backend paths, or deployment internals just to generate an image.
RC2 brings Snapdragon AI Studio significantly closer to that goal.
✨ Major Improvements Since RC1
🧭 Completely Improved First-Run Experience
The first-run experience has been substantially redesigned and stabilized.
RC2 introduces:
-
clearer system and model readiness information
-
improved first-run guidance
-
more reliable clean-install behavior
-
automatic handling of required tokenizer resources
-
bundled CLIP tokenizer release resources
-
improved detection of installed and missing components
-
fewer manual configuration steps
-
clearer guidance when additional model resources are required
The objective is to let Phoenix handle as much of the technical setup as possible.
📦 Beginner-Friendly Model Installation
Model installation is one of the largest areas of improvement in RC2.
Phoenix now provides substantially more guided workflows instead of exposing users directly to model packages and backend internals.
Improvements include:
-
simplified model selection
-
guided installation workflows
-
clearer model names and descriptions
-
improved download, preparation and installation feedback
-
package validation
-
improved installation detection
-
automatic model activation where applicable
-
improved navigation after installation
-
guided handling of externally provided model resources
-
more robust production model paths
Installing a supported model should increasingly feel like using an application — not configuring an AI development environment.
🧠 Stable Diffusion 1.5
The SD1.5 workflow received extensive first-run and clean-install stabilization.
RC2 includes improvements to:
-
clean installation
-
model preparation
-
tokenizer availability
-
production model paths
-
package validation
-
Snapdragon QNN/NPU execution
-
image-generation reliability
The final RC2 production workflow was validated from a clean state through actual image generation.
🧠 Stable Diffusion 2.1
SD2.1 installation and generation have also been significantly stabilized.
RC2 improves:
-
guided model installation
-
model discovery
-
package validation
-
production model-path handling
-
model activation
-
Qualcomm QNN/NPU execution
-
end-to-end generation reliability
A production model-path issue discovered during final RC2 testing was corrected and subsequently validated through real image generation.
🧠 Stable Diffusion 3.5
Stable Diffusion 3.5 represents an important expansion of the Phoenix Engine in RC2.
The SD3.5 workflow uses a Qualcomm QAI AppBuilder-based execution path, extending Phoenix toward newer image-generation architectures on Snapdragon hardware.
RC2 includes:
-
SD3.5 installation workflow
-
guided three-step setup
-
Qualcomm QAI AppBuilder integration
-
improved executable detection
-
correct Python executable handling during setup
-
prevention of duplicate Snapdragon AI Studio launches during setup
-
improved model initialization
-
validated SD3.5 image generation
-
corrected handling of large seed values
-
improved Phoenix Boost compatibility
The final SD3.5 workflow was validated using real installation and generation scenarios.
⚡ Phoenix Boost – An Original Phoenix Engine Innovation
Phoenix Boost was conceived and developed by Holger Kreuzhofen as an original feature of the Phoenix Engine.
Phoenix Boost is designed to enhance image-generation prompts directly inside the local generation workflow rather than turning prompt optimization into a separate cloud service.
It combines:
-
deterministic local prompt enhancement
-
optional local AI-assisted enhancement
-
model-aware prompt optimization
-
prompt preview before generation
-
stable and reproducible seed handling
-
automatic fallback when optional AI components are unavailable
-
integration with SD1.5, SD2.1 and SD3.5 workflows
The core idea is to make prompt enhancement an integrated part of the generation engine while preserving the local-first philosophy of Snapdragon AI Studio.
Phoenix Boost remains optional. Standard image generation does not depend on it.
RC2 substantially improves Phoenix Boost stability and integration, including compatibility with the SD3.5 workflow and corrected handling of large seed values.
To the best of our knowledge, this particular combination of local, integrated, model-aware prompt enhancement and its implementation within the Phoenix Engine is unique to Snapdragon AI Studio.
Phoenix Boost reflects a broader Phoenix Engine principle:
Hide unnecessary technical complexity while giving users more capable local AI tools.
⚙️ Qualcomm QNN & Snapdragon NPU
Snapdragon AI Studio continues to focus on native local AI execution on Windows 11 ARM64 systems powered by Snapdragon X Series processors.
RC2 further stabilizes workflows around:
-
Qualcomm QNN
-
Snapdragon NPU / HTP execution
-
QNN model packages
-
runtime validation
-
model component discovery
-
production model paths
-
backend selection
-
model-specific execution paths
Stable Diffusion 3.5 additionally uses the Qualcomm QAI AppBuilder-based execution path.
🎛️ Generation & ControlNet Improvements
The generation workspace has received additional refinement since RC1.
Improvements include:
-
improved generation controls
-
refined parameter organization
-
improved ControlNet Canny integration
-
improved ControlNet image handling
-
Canny preview workflow improvements
-
improved generation feedback
-
prompt workflow refinements
-
better workspace organization
The goal is to retain advanced functionality while reducing unnecessary interface complexity.
🎨 User Experience Improvements
RC2 includes extensive usability improvements throughout Phoenix.
These include:
-
clearer first-run guidance
-
simplified model selection
-
improved Model Manager workflows
-
clearer installation states
-
improved progress indication
-
better status and error communication
-
cleaner model display names
-
improved generation-status localization
-
improved navigation
-
more consistent dialogs and controls
-
reduced exposure to internal technical details
Users should not need detailed knowledge of QNN, ONNX, tokenizer resources, model packages, or backend architecture to use the application.
🛡️ Stability & Production Readiness
A major part of RC2 development focused on issues that only become visible during real installations and real first-run scenarios.
RC2 improves:
-
installation-root handling
-
model-path resolution
-
model package validation
-
clean-install reliability
-
tokenizer resource availability
-
setup execution
-
executable selection
-
protection against duplicate application launches
-
seed handling
-
backend initialization
-
generation status handling
-
fallback behavior
🧪 Real-World RC2 Validation
The final RC2 build was validated using real production workflows — not only automated tests.
Successfully validated scenarios include:
-
SD1.5 clean installation
-
SD1.5 image generation
-
SD2.1 installation
-
SD2.1 image generation
-
SD2.1 production model-path handling
-
SD3.5 installation
-
SD3.5 image generation
-
SD3.5 setup and launch behavior
-
correct Python executable handling for SD3.5 setup
-
SD3.5 guided setup
-
large-seed handling
-
Phoenix Boost with SD3.5
-
first-run tokenizer handling
-
bundled CLIP tokenizer resources
These end-to-end tests were performed against the production application workflow.
📈 From RC1 to RC2
RC1 established the foundation for Snapdragon AI Studio 2.0.
RC2 turns that foundation into a substantially more integrated and mature product experience.
The biggest improvement is not a single feature.
It is the integration of the complete workflow:
Installation → First Run → Model Selection → Model Installation → Validation → Activation → Image Generation
RC2 represents a major step toward making native local AI image generation on Snapdrag...
Snapdragon AI Studio 2.0 RC1
Snapdragon AI Studio 2.0 RC1
Snapdragon AI Studio is an open-source AI development environment for Windows on ARM, powered by the modular Phoenix Engine.
This release represents the first public release candidate of Snapdragon AI Studio 2.0.
Highlights
- Native Windows 11 ARM64 desktop application
- Designed for Qualcomm Snapdragon platforms
- Modular Phoenix Engine architecture
- ONNX Runtime integration
- Qualcomm QNN Execution Provider support
- CPU Execution Provider support
- Local AI image generation workflows
- Model discovery and backend selection
- Integrated diagnostics and logging
- Modern multilingual desktop interface
- Light and dark theme support
- Standalone Windows installer
Supported Platform
- Windows 11
- ARM64
- Qualcomm Snapdragon X Series recommended
- Python 3.11 ARM64 for source-based development
AI Backends
- ONNX Runtime CPU Execution Provider
- Qualcomm QNN Execution Provider
Backend availability depends on the installed runtime, drivers, models and supported hardware.
Release Status
Snapdragon AI Studio 2.0 RC1 is feature-frozen.
No additional features are planned for RC1. Development following this release will focus on validation, documentation, bug fixes and preparation for RC2.
Validation
The RC1 build completed the project's automated validation suite and was tested for:
- Installation
- Application startup
- Core user interface workflows
- Backend detection
- Model discovery
- Uninstallation
Known Limitations
- AI model compatibility depends on the exported ONNX graph and selected execution provider.
- Some models may not support every available backend.
- CPU-based image generation can require substantial processing time and memory.
- QNN compatibility depends on the Qualcomm AI Stack, device drivers and model configuration.
- Code signing is not currently included in the RC1 installer.
- Windows may therefore display a security warning when launching the installer.
RC2 Roadmap
Planned areas of investigation and improvement include:
- Expanded documentation
- Improved onboarding
- Additional model compatibility
- Performance optimization
- Enhanced diagnostics
- User interface refinements
- Community feedback integration
- Further Qualcomm QNN and Snapdragon NPU evaluation
- Investigation of larger generative AI pipelines
Feedback
Please use GitHub Issues to report reproducible bugs, compatibility problems or documentation issues.
When reporting a problem, include:
- Windows version
- Device and Snapdragon processor
- Selected AI backend
- Model name
- Relevant diagnostic output
- Steps required to reproduce the issue
Disclaimer
Snapdragon AI Studio is an independent open-source project.
Qualcomm, Snapdragon and related product names are trademarks of Qualcomm Incorporated or its subsidiaries. This project is not affiliated with, sponsored by or endorsed by Qualcomm.
Thank you for testing Snapdragon AI Studio 2.0 RC1.