I was curious how Zimage base would behave distilled so it used, the https://huggingface.co/alibaba-pai/Z-Image-Fun-Lora-Distill , one of my trained Loras, and bit of the Muscular women Lora.
Did most of the testing on Euler / simple, cfg 1, 10-12 steps.
Res Multistep or Heun can be a good alternative.
Description
Slightly modified merge with a noticeably different character.
More cinematic presence and expressive realism compared to the base v1.
FAQ
Comments (11)
Do you find that the image variation gets hurt by merging in the 8 step lora? that is what I found in some of my testing.
Seems fine are you using it at 0.7-0.8 strength?
@TiwazM I found I have to put the speed lora down to about 0.3 (and increase the steps) to get some good variation back, and it is still not as good as without it. There is no free lunch I guess, it is fine if you only do a couple images per prompt, I tend to do about 10-15 images and pick the best 1-2 to creative upscale.
@J1B yes true maybe they will improve it. I am still fairly happy with how well it worked. Better than that distilled model someone made…
@J1B I did not realize there are new loras did you try the 2602 versions?
@TiwazM Yes, but I haven't got it to work properly yet, I think there are issues when using it with some schedulers.
@J1B even the first version I thought did not play nice with some combinations
This new model is awesome! Really! The lora 8 stps fusion is good and dark tones almost good! Painting is a regal!👍
Thanks, looking forward to to see more of your pictures:)
@TiwazM 👍
The Diva edition is absolutely fantastic at implementing prompts. In my opinion, it's currently the best checkpoint for that. However, since I prefer realistic images, I need to refine with a different checkpoint. I'm currently downloading the Speed Core, let's see what it can do. I'm looking forward to it :)