3/3/26 UPDATE
I've uploaded a custom merge of the 2603 2-step lora and my key edit of the 2602 8-step lora. This has resulted in a nice, flexible distill that can generate great outputs from 4-10 steps. It has also helped the gridmarking artifacts that were visible at 1.0 strength with the 8-step 2602 UDCAI Edit lora. Please check it out!
Original Post
Credit to alibaba-pai for their fantastic work! This is a step and cfg distill lora to be applied to Z-Image base model and its derivatives. The lora has been converted to be compatible for use in ComfyUI.
Best run at 8 steps with CFG 1. Avoid sageattention. Samplers that add noise per step such as LCM or dpmpp_SDE seem to work well.
Description
New 2-step lora trained by alibaba-pai. The tests posted are using bong_tangent scheduler for consistency across comparisons but I do not recommend it for such low steps. Better to go with something like simple or beta.
FAQ
Comments (7)
Thanks UDCAI.
2603 UDCAI Edit works great for me: 8 steps, cfg 1.5, euler ancestral + simple.
Is it just me or is every noise image pattern start with aliens? LOL I know I sound nuts, but the initial noise patterns are almost all large headed alien silhouettes.
I’ve definitely come across that on occasion, but certainly not every time lol.
@UDCAI Must be the large headed alien porn I'm prompting then... LOL jk
true dude. SO FUNNY
best ZIB-Lora on my SSD! great work, i use it at 0.6 with 12 steps too. and it works so fantastic.
Does those speed up zimage turbo in any form?