New sampler for the Z Image model

[abzaloff]

Hello! Please create a new sampler for the Z Image model. It should run with a low CFG for the first two steps and then raise the CFG to the specified value. Let the first two steps use CFG=0. The model really lacks variability. With such a sampler, we would get random noise at the start of generation and could achieve more diversity from our prompts. This has already been implemented in Comfy by sequentially using two KSamplers.

[Haoming02]

Reference: https://www.reddit.com/r/StableDiffusion/comments/1p94z1y/get_more_variation_across_seeds_with_z_image_turbo/

[Haoming02]

I was able to achieve the effect by using empty prompts during the first few steps.

But interestingly, this method does not work with Qwen-Image. It will simply continue with the blobs it starts with...

[Haoming02]

Trying to develope a more fool-proof method...

[Haoming02]

Ladies and gentlemen, this has been implemented~

Enable it in Settings/Stable Diffusion

2 steps

[abzaloff]

Thank you so much for your hard work!!! Low bow!

[DenOfEquity]

prompt editing: [starting prompt, maybe empty : real prompt : 0.125]

only seed changed, empty 1st prompt. Except 1st image which also has a photographic style applied that tends to start dark.

[Haoming02]

(Note: this wouldn't work in Neo as I did not pad the conds XD)