Z-Image Lora training test, Used for Z-Image aesthetic enhancement, improving illustration and artistic concepts.
Z image turbo v1: Recommended settings: heunpp2 + linear_quadratic, cfg: 1.5, step: 20, 0.8-1.2
Z image Base v1: euler_ancesral + beta，30 step，4 cfg，0.6-1 weight
Try it online: Tensor | Seaart | Liblib
Z image turbo v1: The LoRa dataset uses a variety of training sets, which affects various styles, including realism, illustration, and animation. It also fixes the aesthetic issues of Z-image. Because LoRa training disrupts the acceleration effect, it requires 20 steps and cfg 1.5-2.0 to be effective. You can try the others yourself.
Due to the difficulty of Z-Image training, the step count and CFG need to be increased. I will update this when a better training method is available in future Base versions. Have fun!
Z image Base v1: Z image Base v1 has been updated; other versions are only released on TensorArt.
we sincerely apologize to users who previously downloaded the z-image base v1 version. as there was no Zimagebase category on Civita at the time of the lora's release, it was mistakenly labeled as Zimgturbo. please try using the base version. or try it on tensorart online
Description
For Z-image BASE
euler_ancesral + beta，30 step，4 cfg，0.6-1 weight
FAQ
Comments (9)
This lora is so amazing.thanks
https://civitai.com/images/119623129 - How did you achieve the image in the style of illustrious using only the basic model and this lora? By applying this lora, I got something closer to realism, and without lora, something on the contrary is more similar to what you showed.
The style and details of this image come from another Lora, which hasn’t been released on Civita as it still requires adjustments.
If you're interested, you can try it online at Tensor.
@Ray3780 Thanks for the reply and the link, but I prefer local generation and experimentation. For the last week, I've been trying out a variety of datasets, workout settings, and tags on Z-image turbo + Z-image Base without a break. So far, my observations of the training and generation process have led me to abandon the turbo version as a base for lora training and use it only for pampering and quick generation as it is, since lora ZiT and ZiB are not particularly friendly. If you train lora on ZiT (I focus on anime in the illustrious style), then the further you train, the more the anatomy of the characters breaks for some reason. It manifests itself in broken fingers and porridge instead of legs, despite the fact that they do not appear in the dataset and tags for training at all.
I apologize in advance if something was badly translated, not my native language + translator
@Sansenskiy ZIB is better suited for training anime styles, while ZIT is more suitable for photorealistic style training. ZIT offers two training methods. I only master bypass distillation training. The drawback of this method is that it undermines ZIT's acceleration performance and requires more training steps, while its advantage lies in a higher learning rate.
This is actually insane, it makes ZIMG Base go from unusable to easily the best model I've ever used.
Thanks. ZIB is more versatile than ZIT, and the speed isn't actually that much slower. There's a bug with ZIB's LoRa training, but I've still trained seven ZIB LoRa instances. The rest are in Tensorart. I think the Art shoot and Unlock XXX are good; if you're interested, you can help me test them on Tensor. Once the version is perfected, I'll put it on Civita.
should i use it with SDA?