The purpose of this finetune is to improve the visual aesthetics of the Z-Image "Base" (Non-Omni) model. This is very much a proof of concept finetune. Its beta stage and will likely need more refinement.

Update as of 1:35pm PST: I just fixed the file. It should work in ComfyUI now. It is called ZImage_Eminence_v1_fix.safetensors

Version: Beta 1

Features

• Improved visual aesthetics

• Maintains base model's general results

• Trained on word tags, short and long descriptions to maintain broad understanding. (Z-Image paper: https://arxiv.org/pdf/2511.22699)

Notes

• All images are oneshot generations using the same seed. They were generated using musubi-tuner.

Issues

Unknown. Sampling looks decent across a variety of topics but I am not aware of any systematic issues but I'm sure you will let me know.

Shoutout

OneTrainer: Awesome job everyone! Seriously, if you are not using OneTrainer, I would definitely recommend checking it out. Lots of functionality and they offer features/concepts that are taken directly from research papers.

Musubi-tuner: kohya-ss is a legend! Another great trainer. Updated quickly and great notes.

Z-Image Team: Thank you so much for the detailed paper. A great source of information, strategy and concepts. I can read 100% of it and understand about 15% of it but the 15% is enough to emulate concepts at-home. I would highly recommended checking out the research papers. These are a great source of information.

Civitai: Thank you all the hard work and posting/updating everything so quickly!

ComfyUI Team: Thank you for the awesome community support and continued development!