Add basic support for Z-Image Turbo txt2img

[monstruosoft]

This PR adds basic support for Z-Image Turbo txt2img mode by making use of the Z-Image-Turbo-SDNQ-uint4-svd-r32 quantized version of the model. RAM usage is ~9 GB when using this SDNQ quantized model, other SDNQ versions of the model might work as well but I have only tried this one.

In order to run Z-Image Turbo, it's necessary to install the sdnq python package and upgrade several packages, including upgrading diffusers to the latest source version. I didn't update the requirements.txt file in case this PR waits for the Z-Image support to make it to the official release of the diffusers library.

[rupeshs]

Seems like it needs more inference steps (around 9?)

[monstruosoft]

I've got some decent results with as low as 4 steps, but yes, 8 to 10 is the recommended number of steps.

[rupeshs]

Other than sdnq,diffusers what all other packages updated for this to work? did you disable openvino?

[monstruosoft]

The diffusers and sdnq packages have to be installed from source since the ZImagePipeline is not yet in the official release of diffusers:

# pip install git+https://github.com/Disty0/sdnq
# pip uninstall diffusers
# pip install git+https://github.com/huggingface/diffusers

Other than that, I updated the peft, transformers and optimum-intel packages to avoid conflicts. Updating torch and torchvision may or may not be necessary but I updated those anyway. These are the versions of those packages that I have currently installed:

diffusers                 0.36.0.dev0
optimum                   2.0.0
optimum-intel             1.26.1
optimum-onnx              0.0.3
peft                      0.18.0
sdnq                      0.1.2
tokenizers                0.21.4
torch                     2.9.1+cpu
torchvision               0.24.1+cpu
transformers              4.55.4

These are all the packages I updated, didn't touch the openvino package.

[monstruosoft]

Sorry, I was so carried away with Z-Image that I didn't notice that I broke generation with other models. Latest commit should fix it. I apologize for causing any issues.

[rupeshs]

@monstruosoft It seems slow, what is the observed latency for 3 to 4 steps 512 x 512 image?

[monstruosoft]

I wanted to try it offline but it takes almost 10 minutes to generate a 512x512 image on my machine, but my machine is over a decade old, I guess it might still be usable on more recent hardware, even using CPU only.
So, yeah, it is slower than SDXL but I'm guessing it should still be faster than FLUX, while consuming much lower RAM. I mean, I can't even try to run FLUX on my machine.