Generative AI just got better – and faster – on AMD Radeon graphics cards and Ryzen AI processors. Through an engineering collaboration with Stability AI, AMD is proud to announce AMD optimized models that speed up inference on compatible hardware by up to 4.3x.
Read more about the AMD optimized models in this technical deep dive by AMD Fellow Hisham Chowdhury and team: Xiaoxiao Liu, Sonbol Yazdanbakhsh, Lucas Neves and Owen Zhang. For developers, these models can be identified with the “_amdgpu” suffix on Hugging Face.
Users can experience these models and several other exciting new features by downloading the newly released Amuse 3.0 Beta software by Tensorstack.AI.
- Amuse 3.0 and AMD optimized models require the Adrenalin 24.30.31.05 AMD Optimized and Amuse 3.0 Preview Driver (Download Here) or the upcoming Adrenalin 24.5.1 mainline release.
What’s new in Amuse 3.0:
- Introducing AMD optimized models (created in collaboration with Stability AI)
- Introducing high quality Photo AI Filters
- Introducing draft quality Video Diffusion
- Introducing draft quality Video Restyle
- 100+ new image models and fine tunes added
Amuse 3.0 allows AMD users with compatible hardware to run state of the art models like the Stable Diffusion 3.5 family by Stability AI and FLUX.1 models by Black Forest Labs. 100 of the most popular custom image models and fine tunes were also added in this release to give users a one-stop shop to maximize their creative potential.
Users with AMD Ryzen AI systems powered by an AMD XDNA NPU can use the AMD XDNA Super Resolution feature to easily and quickly create 4MP print-quality images in a matter of minutes.
Recommended system requirements
| Models | AMD Radeon Graphics Cards | AMD Ryzen AI Processors |
| SD 1.5 and LCM | Most AMD Radeon Graphics Card with at least 6GB of vRAM | Most AMD laptops with 16GB RAM. |
| SDXL, SDXL Lightning, SD3 Medium, SD 3.5 Medium | AMD Radeon RX 9070 XT, Radeon 7900 XTX, Radeon 7900 XT, Radeon 7900 GRE, Radeon 7800 XT | AMD Ryzen AI 300 series with 32GB RAM, Ryzen AI 395+ MAX series with 32GB RAM. (VGM = 16GB) |
| FLUX.Schnell, SD 3.5 Large Turbo | AMD Radeon RX 9070 XT, Radeon 7900 XTX, Radeon 7900 XT, Radeon 7900 GRE, Radeon 7800 XT | AMD Ryzen AI MAX+ 395 series with 64GB or 128GB RAM. (VGM = 48GB) |
| SD 3.5 Large, FLUX.Dev | AMD Radeon PRO W7900 XT 48GB, Radeon PRO W7800 48GB |
The integration of AMD optimized models inside Amuse allows users to get an inference speed up of up to 4.3x depending on the hardware used. On the AMD Radeon RX 9070 XT, users can expect up to 4.3x faster image generation in Stable Diffusion 1.5 and 3.1x faster performance in SDXL 1.0. In the SD 3.5 family of models, AMD optimized models deliver up to 3.3x faster performance in SD 3.5 Large on the AMD Radeon RX 9070 XT.
Users with AMD Ryzen AI 9 HX 370 will see speed ups between 5% and 50% depending on the model chosen as well as being able to run larger models than ever before. Running SD 3.5 Large Turbo for example, is only possible with AMD optimized models. AI enthusiasts with AMD Ryzen AI MAX+ 395 (codename: Strix Halo) series processors will experience up to a 3.3x speedup in SD 3.5 family of models and up to 70% and 40% faster performance in Stable Diffusion 1.5 and SDXL 1.0 respectively.
Of course, AMD optimized models will offer inference speedups on most modern AMD graphics cards and processors, but mileage will vary depending on the exact configuration, architecture, family and model chosen.
Amuse 3.0 isn’t all about performance though. We heard that you like high quality photo AI filters. The latest version of Amuse combines high quality SDXL Lightning based models with control nets to give you stunning Photo AI filters – running completely locally. Want to hop on the latest AI image trend and don’t want to put your picture into a cloud service? Run Photo AI filters completely locally and privately on your device.
It’s not just photos that are getting all the love this time. Video Diffusion support is now officially available in Amuse 3.0. Re-styling your videos with the draft quality video-to-video model or text-to-image generation of clips is now possible entirely locally.
