I really like DMD2 - not because of its speed. But because with DMD SDXL improves distortions caused by VAE, details like small objects are much more coherent.

Maybe a DMD version for SDXL trained in 8 steps would be even better. 

And I believe it is worth training DMD with SD 3.5 and Flux. 

perhaps DMD would make it possible to generate coherent text from SDXL. I don't know if it's possible to train a mini Lora DMD in specific subjects like text