Please leave a rating if you enjoyed this LoRA! Its Free!
Ill Write it 3 times because its important. ONLY USE THIS WITH A DISTILL MODEL
CarlesDalmau LoRA
This LoRA is designed to imitate CarlesDalmau's unique style. It can produce drawings as he draws them, or stylize more human characters. Tagging was done by Qwen3 VL 32b.
What do I need to know?:
This Lora is Trained on ZiB, but is intended to be used on a distill, I personally use THIS ONE.
This LoRA is Trained on Natural Language Tags, using booru tags will be ineffective for getting the results you want. Look at my example photos for the basics. Works with both long and short tags.
The recommended Weight is 1,
The model works well at 1024, but works even better at 2048
If you are using the Distill I recommended, use cfg 1.5, and shift ~6-7, with sampler euler simple
This Lora basically requires the word "Digital Illustration" to work. Its not in my dataset, dont ask me why.
Again, This Model is NOT Intended to work with ZiB, only a ZiB Distill (But not ZiT)
Thank you!!
Description
FAQ
Comments (4)
Glad to see Z-Image Base takes on art styles as well as Z-Image Turbo. It's honestly the most surprising thing about Z-Image, is that a model SO focused on photo portraits, ended up taking on art styles better than most models.
works better than Klein 9b in my experience tbh. But it HEAVILY relies on keywords and models. This LoRA straight up wont work without prompting for "digital Illustration" and using a distill. Just seems to be an idiosyncrasy with the model.
Might try some concepts soon, after I'm done all the style datasets I've prepared
@Erebussy Ah that's a shame. I like to stay away from triggers for styles, so that I can mix them easier. Once you have like 3 triggers, and you're brining the strength of one of the loras down, you start getting weird artifacts from the triggers it doesn't fully understand. One of my favorite style loras for Qwen has a trigger that's something like tbear_ink, and it starts putting bears in the image when I lower the strength. I think your use of "digital illustration" as a generic trigger is a good middle-ground for allowing the mixture of styles without artifacts.
@Jellai Interestingly, id say my actual style trigger for my Lora's does approximately 5% of the work. The LoRA works without the "carlesdalmau" trigger more or less perfectly fine. What's interesting is that "Digital Illustration" was completely removed from my dataset, the model requires that in spite of the fact I never tagged for it. So I can only assume that current z-image models are so realism biased that without directly prompting for digital art styles the lora cant possibly overcome the bias.
The good news is, if anyone ever actually does it, finetunes would in theory completely eliminate the need for this, and you could likely drop one or both of the triggers. Whether or not people actually finetune z-image is still up in the air unfortunately