Buuta/dreamshaper-xl-lightning-for-Snapdragon-X-Elite

DreamShaper XL Lightning (Pre-compiled for Snapdragon X Elite NPU)

This repository provides a pre-compiled, optimized version of DreamShaper XL Lightning running in fp16 precision, specifically designed to run 100% locally on the Snapdragon X Elite NPU via ONNX Runtime with the QNN Execution Provider (QNN EP).

To bypass hardware compiler constraints and memory limits, the core UNet architecture has been meticulously split into 5 distinct sub-models.

---

🚀 Key Features

• 100% NPU Execution: Unlike other hybrid approaches, every single component—including both Text Encoders, the 5-split UNet, and the VAE—runs natively on the Hexagon NPU.
• Bypassed Size Ceilings: The UNet is split into 5 fragments to stay under the 2 GB Google Protobuf.
• Sequential Load-on-Demand Pipeline: To manage the strict memory overhead on ARM Windows, the script utilizes an aggressive memory management strategy:
  1. Loads Text_Encoder & Text_Encoder_2 into RAM/NPU ➡️ Processes ➡️ Flushes from memory.
  2. Loads the 5 UNet parts into RAM/NPU ➡️ Processes ➡️ Flushes from memory.
  3. Loads VAE into RAM/NPU ➡️ Decodes the final image.
• Performance: Generates 1024x1024 images around 62 to 68 seconds (with 6 steps / Guidance Scale 2.0).

---

🛠️ How to Use & Implementation Details

All the necessary Python execution scripts, environment setup requirements (requirements.txt), the original UNet splitting pipeline, and step-by-step instructions are hosted on GitHub.

Since this project requires specific pipeline orchestration to sequentially load and flush the models, please refer to our GitHub repository for the full guide and inference scripts:

👉 GitHub: buuta-buta-butaata/SDXL-with-Snapdragon-X-Elite-NPU

---

👥 Credits & Disclaimer

• Base Model: Huge thanks to Lykon for the amazing DreamShaper XL Lightning.
• Project Context: This repository is a Proof of Concept (PoC) co-created by a human developer and Google Gemini (AI collaborator).
• Safety Notice: This model card does not bundle a safety checker/NSFW filter. Please use the generation outputs responsibly.