Stable Diffusion gets up to 3.8x faster on AMD chips

Running AI models locally on AMD hardware has notoriously been a headache compared to Nvidia's plug-and-play CUDA ecosystem. Stability AI is stepping in to change that narrative.

The AI image generation giant just partnered with AMD to roll out officially optimized versions of its Stable Diffusion models, specifically tuned for AMD Radeon GPUs and Ryzen AI APUs.

Up to nearly four times faster

The update targets four key models in Stability AI's lineup: Stable Diffusion 3.5 Large, SD 3.5 Large Turbo, SDXL 1.0, and SDXL Turbo.

Rather than forcing users to fiddle with complex drivers, these builds rely on ONNX Runtime to bypass typical performance bottlenecks on AMD hardware.

The speedups are significant. Compared to running the standard PyTorch base models, the optimized SD 3.5 family delivers up to 2.6x faster inference.

Older favorites see an even bigger jump. Both SDXL 1.0 and SDXL Turbo achieve up to 3.8x faster inference speeds on compatible AMD chips.

Stability AI confirmed that these performance gains come without compromising visual quality, and all the models retain their original open licenses.

Two paths to try it out

For developers, the models are already live on Hugging Face with an amdgpu suffix, ready to drop directly into existing ONNX Runtime workflows.

Casual users who just want to make art without touching a terminal don't have to miss out either.

Stability AI pointed users toward Amuse 3.0, a desktop interface that natively supports these newly optimized AMD models straight out of the box.

Why it matters

Nvidia has held a practical monopoly on consumer-grade local AI inference for years, largely because alternative hardware lacked out-of-the-box software support.

By bringing top-tier models like SD 3.5 and SDXL to Radeon graphics cards and Ryzen AI APUs, this update makes local image generation viable for millions of laptops, desktop PCs, and handheld gaming devices that never had access to decent inference speeds.

In our view, AMD's decision to partner directly with model creators is the right playbook. Competing on raw silicon specs is useless if developers are stranded in driver purgatory.

The lingering question is community tooling. While standalone generation is now fast and clean, it remains to be seen how well these ONNX-wrapped models will mesh with community-favorite tools like LoRA fine-tunes and ControlNet workflows.