Update: V2_DMD2 is a V2 with the DMD2 Speed LoRA included + a custom LoRA to fix contrast and hands. You can use it without the Speed LoRA:
Lower your CFG to 1.2~1.4 (for original effect), ~2.5 for what i feel is a nice result.
Switch to Euler A+Auto or LCM+Normal
~8 steps (both for normal and hires)
Small update: Thanks to clueless_engineer i realized that this model can do GREAT stuff leveraging the DMD2 speed LoRA with low step count! Check the pinned pictures!
This model is an attempt at "fixing" just enough the base Illustrious models to help me training/tests LoRA without too much influence.
To do so, i have done a merge of Illustrious + one of my models at 0.15, CosineB calculation mode.
v0.1 is Illustrious v0.1 + AnBan AlterV3
v2.0 is Illustrious v2.0 stable + AnBan ShinV1
It's really more of a personal test than a real "let's all do this now" 😅
Description
This version includes the DMD2 LoRA and some helpers to keep hands and stuff in check.
=> switch to Euler A + 8 steps + CFG ~1.5 - 2.5
FAQ
Comments (2)
CFG 3 might be too much with the DMD2 lora. I run at 1.2 to avoid too much contrast, at 8 steps. It works great for anime and realistic images.
I did a few generations at 1.4, 2 and 3, it can really depends if you need your negative prompt to be more enforced. 😊