This is a merge between some illustrious models i like.
Personally i like to use DMD2 (4-Step) Lora for speedy generation, so i baked it into the model with version 3.
Settings
Version 3 - DMD2
Make sure to only use 4-8 Steps
LCM Sampling
CFG Scale 1-1.5
You can play with Schedulers, but Karras or Align Your Steps works best
Version 1 and 2
For non DMD Version (1 and 2) use any Settings you like:
CFG: around 4-5
Sampling: Euler a or 2M SDE
Scheduler: SGM Uniform
15 steps +
Description
v3.2 contains more merges too many to list, the last being CyberIllustrious, slightly more realistic.
Some seeds in the comparison images are changed.
== Version 3.x is a DMD2 Checkpoint ==
Make sure to only use 4-12 Steps
LCM Sampling
CFG Scale 1-2
You can play with Schedulers, i prefer Align Your Steps
FAQ
Comments (6)
V3.2 is a very nice semi-realistic Illu, super fast, and absolutely beautiful images. I use it with Euler a/Karras 7 Steps, 1.0 cfg. With these settings, I get nice colors, good contrast, and good details. Thanks for this checkpoint.
Cannot get anything remotely coherent out of 3.2... Not even when blindly copying all generation data from the sample images. Does this require a specia VAE? I get greyed out, unsharp images that don't follow the prompt at all. (Local generation, SD/A1111)
hm no, are you sure to copy settings? LCM Sampling,1.5 CFG and 10 steps ?
I'm using forge, i am not sure if A111 is still being developed or if it stopped completely by now.
Development in both has slowed to a crawl, but forge is a bit more up to date,
@doombatz Yes, I have tried several, copied everythign exacly. Same greyed out ghosted results. In the meantime I did some searchiing and read somewhere that this can happen in A1111 if you use an LCM pipeline because A1111 isn't optimized/coded for that. The closest I got so far was by upping the CFG to 4.5 and steps to 30+ and using ++2M/Karras. But the the result was grainy and contrasty.