I would like to do a full fine tune of Qwen-Image using diffusers/accelerate, similar to the SDXL training process laid out here.

I am not looking to do Dreambooth or a LoRA. I have a large (~5 million image) dataset that I would like to use to do a full-parameter fine tune of the model. I am aware that other libraries such as kohya-ss/musubi-tuner have experimental support for Qwen-Image fine tunes and LoRA, but I am curious about learning to do it with only HF libraries.

Is there a guide anywhere for how to do this with Qwen-Image? Can someone provide me with an example command for training Qwen-Image?

Thanks!