Unlock diversity of Z-image-Turbo, comparison

Quick Comparison:

1. Skip steps and raise the shift to unlock diversity of Z-image-Turbo
2. Significantly increase Z-Image Turbo output diversity with one ComfyUI node
3. Z-image diversity from Civitai entropy

Image variations without cherry picking :

1-A

https://preview.redd.it/73c38zw6z35g1.png?width=1062&format=png&auto=webp&s=d1a7e040c0d7c710bb574d3070fefb444c1e8336

1-B

https://preview.redd.it/nox8u17u035g1.png?width=1064&format=png&auto=webp&s=55385d05a148e2bb52595c3e6ffa1db983252bf6

3-A

https://preview.redd.it/620pj2vu035g1.png?width=1067&format=png&auto=webp&s=35f629749b0b559cbe09016e3500e7064d82bd31

3-B

https://preview.redd.it/x8eswqlv035g1.png?width=1064&format=png&auto=webp&s=6aae296da4ee3d292659759ba3989054d65ca155

3-C

https://preview.redd.it/pdtgeyf3335g1.png?width=1041&format=png&auto=webp&s=224c904f6d15951cf2967868269e5b765bfa0cae

2-C

https://preview.redd.it/pl73rww4335g1.png?width=822&format=png&auto=webp&s=5609264eff447a175b346a09d43946269db3fa49

Observation:

Z-image's randomness from empty steps tends to generate human portraits, so method 1 seems OK for PROMPT A but not really works for PROMPT B.

While providing random photos with diverse palette/light/shadow works for both cases.

====
Updated for 1-A: 4 photos in similar batch in red rectangles should be removed from variety comparison. MY BAD!

https://redd.it/1pdluxx
@rStableDiffusion

The prompt adherence of Z-Image is unreal, I can't believe this runs so quickly on a measly 3060.
https://redd.it/1pdsz9x
@rStableDiffusion

Z image inpainting with new support from LanPaint
https://redd.it/1pdsyo5
@rStableDiffusion

Z-Image styles: 70 examples of how much can be done with just prompting.

https://redd.it/1pdy78q
@rStableDiffusion

Z-Image Only on CPU+16GB Ram some images i generate on it

https://redd.it/1pdyvpi
@rStableDiffusion