Advancing AI with Nexa AI: Image Generation on AMD NPU with SDXL-Turbo

Nov 05, 2025

AMD continues to push the boundaries of on-device AI acceleration through strategic collaborations that unlock new possibilities for developers and researchers. Today, we're excited to highlight ongoing work with Nexa AI, bringing cutting-edge capabilities to the AMD hardware ecosystem: SDXL-Turbo image generation on AMD Neural Processing Units (NPUs) and Gemma-3n multimodal inference on AMD integrated GPUs (iGPUs) through the NexaML engine.

Nexa AI is an on-device AI deployment and research company with the goal of making on-device AI friction-free and production-ready. The collaboration was prominently demonstrated at PyTorch Conference 2025, showcasing the technical achievements made possible through the joint engineering efforts of AMD and Nexa AI.

Photo of the AMD Booth at PyTorch Conference 2025 showing the image generation demo on Ryzen AI 395+ Framework Desktop

SDXL-Turbo: Real-Time Image Generation on AMD NPUs

SDXL-Turbo represents a significant breakthrough in text-to-image generation technology. As an advanced distilled version of Stable Diffusion XL (SDXL) 1.0, it has been specifically optimized for real-time generation workflows. Through AMD and Nexa AI's collaborative optimization work, this model now leverages AMD NPUs to deliver high-quality images.

Running SDXL-Turbo on AMD NPUs takes just two simple steps with NexaSDK:

Download Nexa SDK from the Nexa SDK model hub.
Execute with a single command:

    nexa infer NexaAI/sdxl-turbo-amd-npu --prompt "a cartoon cat"

Once the SDXL-Turbo model is loaded into memory, image generation begins—executed entirely on-device with no cloud dependency, and accelerated by the AMD Ryzen AI NPU.

Developer-Friendly Features with NexaSDK

NexaSDK provides additional developer-friendly features to help you build applications faster. The SDK offers Python bindings for seamless integration into your applications, as well as server mode for rapid prototyping with OpenAI-compatible APIs. For more details, see the Nexa AI documentation on Python bindings for Windows x86_64 and server mode.