UnicomAI/LeMiCa — reverse-engineered prompt

Build me a working LeMiCa setup that I can actually use to speed up diffusion image and video generation.

I want a simple way to run supported models through LeMiCa and get noticeably faster results without training anything. It should feel practical, not just like a research repo. Please include the main supported options from this project, like ComfyUI use, plus the model folders here such as Wan 2.1, HunyuanVideo 1.5, Qwen Image, FLUX, ERNIE Image, and Z Image, wherever the repo already supports them. I also want easy speed presets like slow, medium, fast, and ultra when available, so I can choose between quality and speed.

Please make the setup clear enough that I can install it, pick a model, enter a prompt, and generate images or videos with LeMiCa enabled. Add a short README with exact run steps, basic examples, and any model specific notes. If anything is missing in the repo, look up the current docs online and fill in the gaps.