OOXML is incredibly complicated, but LLMs can write it well enough. One of my colleagues was using an LLM to generate PNG slides for a presentation. They looked pretty, but changing anything was a pain. I fed the images to free ChatGPT and asked for a PPTX with every text box, frame and image split into its own object, and it spat out a real, editable presentation.
You could probably then copy-and-paste those into a PPTX file created by PowerPoint to make sure there were no weird things going on with the file(structure) that might cause problems later.