DISCLAIMER
This article only concerns creators who distribute SDXL family checkpoints patched with DMD2 only.
This article is only valid if I understand how to use DMD2. Otherwise please give me your magic settings in the comments (other than CFG=1.00001 Step=12)
So...
These days, I see more and more checkpoints downloadable only with integrated DMD. This therefore means:
very fast image generation
LCM sampler, with almost no choice for scheduler
ridiculous CFG (or =1, therefore without negative prompt)
But it's not always good.
This thing is supposed to save us time, except that if I want to play a little bit with the CFG, I have to do lots of tests so as not to end up with disgusting textures and/or sick contrast.
I also read that it gives identical faces.
DMD2 is great, I use it especially for the second pass (Hires Fix), but full of limitations. Remember that SDXL does not understand prompts like Flux (which is not ideal), so any change to its understanding has a big impact.
I no longer download checkpoints with DMD because we can all use the DMD2 LoRa on a "normal" checkpoint, in a prompt or by doing a merge on a normal checkpoint if we want.
Thank you for giving us the choice.
Free will is good
(remember and apply this here and for everything you do in your life).
Thank you !
PS : I wanted to make a bad burned picture with a DMD2 model, LCM, CFG=4 and 20 steps for the header... But the result is better than if I actually wanted to make it like that... As always.
PPS : For me, DMD2 is a "usefull destructive" option that does not reflect all qualities of a checkpoint. Like MP3 vs WAV : you don't see the difference until you do.