aww thanks so much!! I really needed this Lora to fix ZIT, so that I can continue making stuff in my preferred style, so here we are =) Looking forward to seeing what you create too!

The Z Image standard images are very "raw", and I feel like lacking in a bit of beauty composition, unlike the better balanced Qwen model (in particular White Qwen checkpoint). This lora at lowish strength will help bring some of that in, and bring back a bit more of that cinematic vibe. If you go too high on lora strength though it would take away too much original Z image, however, using a chained sampler setup could help to bring back more of the Z Image detailing.

@zengrath @letthebassdrop That's interesting, because I don't have any trouble with having varied faces, I generated around 200 last night, and 90%+ of them are different faces. I do suggest not going that high on the lora strength, otherwise it will overpower the original Z Image.

thanks! I heard that there's a cap on how much lora weight can be mixed right now. I think my lora is pretty gentle though, so should be fine to mix but I haven't tried. I assume when mixing loras etc, to have one be dominant and another be less, instead of 50/50.

@EarNugget2 I think the sweet spot for this lora is between 0.2 to 0.5, but you can go even higher if you don't mind an overly smoothed look, really depends on the image you are going for! I didn't notice any bad anatomy even at 1.0 strength, but there are some weird artifacts like random tattoos =)

I can't afford to train it that many times, so I purposely used R128 to try to get in as much detail-adding features as possible.

ai toolkit, I used r128, and 5000 steps. (since it's realistic content so I needed higher amounts.) It's pretty much default settings, just have to make sure images and captions are good.

aww, I dunno if it's as good as Midjourney, they're probably much better than my images, but thank you for saying that, will keep trying to improve!! Actually I found new improvements just now, uploading new pics of super crisp fashion style.

@aneebtheartist126 let me think about what would work best with this, there are so many ways to upscale. What did you need upscaler for? It's fast enough to generate at higher resolution already if you want, and Z image doesn't need as high resolution to get good details when using my workflow.

@aneebtheartist126 ok I worked out the best way to do upscale without the doubling faces issue, once I do some final tests I'll push it as a workflow version update!

It's a feature. That means she'll never lose her sword ever again.

aww thank you for using it!! I hope to see more and more of the artwork that you create with my lora(s)! =)

Thank you so much, I put the workflow together here instead of separate since it's so essential for making Z Image look good (weird that Z image by default has a noise bug) , and wanted to make sure my lora's users get the best possible opportunity to make great pics!! I really appreciate your kind words and you sharing your artwork here =)

@DiabolicX I have lora#2 in training, as soon as Civ actually processes it instead of cancelling it. And I'm building the dataset for #3. Each of them should be very unique visually and fun to play with! =)