This is my version of Z Image Turbo.
The goal was to incorporate better colors, more contrast, and above all, more variety into the checkpoint. You should decide for yourselves whether I succeeded.
To achieve my goal, I created various LoRAs and then incorporated them into the checkpoint.
Test this checkpoint, post your results directly here on the model map, and please leave a comment. Constructive criticism is welcome. Perhaps it will lead to further improvements.
The checkpoint works with all samplers that you know from the Z Image Turbo model.
I recommend "DPM++ 2s a RF" as the sampling method and "Beta" as the schedule type. Even though this sampler is the slowest and therefore doubles the calculation time, it achieves the most variations with excellent quality.
Update V3:
The model file size is now only half that of V2, yet:
- improved skin detail
- more realistic lighting
- more realistic shadows
Even though good results can be achieved with 9 or 10 steps, I usually process most of my images with 30 steps to fully utilize the checkpoint's potential.
Sampling Method "Euler a", Schedule Type "Beta"
If you like the model, I would be very happy to receive a "thumbs up."
Creating a checkpoint is very time-consuming. You can support me with a few BUZZ.
Description
FAQ
Comments (6)
This website can no longer display whether the file is of the FP8, FP16 or BF16 type! The website is becoming increasingly stupid!
@Mr_Fei They made a change where the creator can chose from fp16, fp8 , nf4 , q8 , q6 etc , try to find my model and see if you can see what it is
@BastardOG bf16
@BastardOG There was no indication at all. I couldn't figure it out!
@Mr_Fei when there are 2 or more variants it seems .
@BastardOG It seems to be due to the new look of the model card. When uploading the model, I specified that it was BF16. Most Z-Image Turbo models are created in BF16 or BF8, as it's more stable and more efficient on newer NVIDIA graphics cards (RTX 30xx, 40xx).