MiniMax-H3-Image for AMD Ryzen AI NPU
Image Generation on AMD Processors
2 min readupdated
This unofficial port of the MiniMax-H3-Image model is adapted to run on AMD Ryzen AI processors featuring NPU support (XDNA 2). The model lets you generate 1024×1024 images in six steps right on your PC's hardware accelerator. The file set includes optimized weights for the diffusion transformer and decoder. This solution will be useful for owners of compatible AMD Ryzen AI processors running Windows 11.
Key features
- Running the diffusion transformer and decoder on the NPU (XDNA 2)
- Total model file size is 21.81 GB using int4 quantization
- Using Ryzen AI Software 1.8.0 for graph conversion and execution
- Text encoder and token refiner run on the CPU
What you can do
- Generate images from a text prompt on a PC with an AMD Ryzen AI processor
- Set up the environment using GitHub repository scripts and run generation locally
- Test image generation speed on hardware without a discrete graphics card
Example
git clone [https://github.com/tomhda/minimax-h3-image-ryzen-ai-npu](https://github.com/tomhda/minimax-h3-image-ryzen-ai-npu)
cd minimax-h3-image-ryzen-ai-npu
pwsh -NoProfile -File .\setup.ps1
pwsh -NoProfile -File .\doctor.ps1
pwsh -NoProfile -File .\gen.ps1 -Items .\examples\one.json -Out .\outputs\firstRead more
DetailsBenchmarks: on Ryzen AI 7 PRO 350, generating a 1024×1024 image (6 steps) takes 368 s (including 105 s for session creation).
Hardware requirements: Windows 11 on AMD Ryzen AI with XDNA 2 NPU and Ryzen AI Software 1.8.0.
Limitations: the MiniMax-H3-Image Community License Agreement excludes several regions (EU, UK, South Korea, US).














Comments