Dead men tell no tales: Local image AI Amuse 3.2.0 released
Development on the Amuse image and video AI was discontinued, and the sources were released as open source. Now there are initial updates.
(Image: heise medien)
The development of the locally operating image and video AI software Amuse had originally been discontinued by manufacturer TensorStack-AI and the sources published. Now, the open-source project released its first update on Friday night this week. Development is thus continuing.
There have been some changes in the Amuse GitHub project since December. The focus on and strong optimization for AMD hardware is a thing of the past. The project now particularly highlights HP AI devices in its advertising. However, concrete adaptations to the manufacturer's hardware are not apparent.
Not all changes logged
The official changelog for the current version Amuse 3.2.0 appears concise at first glance. The migration of all components to .NET 10 is said to provide a performance boost due to better memory management and JIT compiler optimizations. The update to OnnxRuntime 1.23.0 is said to lead to greater stability and provide support for integrating newer AI model architectures. The execution providers are also said to be improved and deliver more speed. The dependencies have been modernized and the internal NuGet libraries updated to the latest stable versions. This ensures greater security, stability, and compatibility.
Videos by heise
However, in a brief test with a rather underpowered AMD Ryzen 7640HS system with 64 GB of RAM, which only contains a 10-TOp/s NPU, the results are mixed. The changelog does not mention the changes to the user interface; for example, the displays for processor, GPU, and NPU, as well as their respective memory usage, have been scaled back. AMD optimizations such as the AI-based and fast "AMD XDNA Super Resolution" upscaler have been removed.
The AMD AI-optimized models are no longer found in the software. However, the general models function at largely similar speeds as before. However, the OnnxRuntime now crashes more often with models that require more memory, with the error message that the graphics card is no longer accepting commands. The smaller models up to about 4 GB, however, work without problems.
It remains to be seen whether further development will address and resolve these issues. Local Artificial Intelligence offers more privacy, and Amuse is currently the simplest way to locally create images and video clips. Solutions like ComfyUI are an alternative, but significantly more difficult to use.
(dmk)
