The main goal is to improve the look in photorealistic images. Mostly a western influencer style, but it works for other things too.
This LoRA was trained on over 750 high-quality images.
Trained with the de-distilled Z-Image using AItoolkit.
I include the LoRA at 13k, might add more if people want it.
Optimal weight is generally 0.5–0.75, though lower strengths can produce interesting results on subjects outside its main domain.
Captioned with joyCaption v2.
Thanks to Ostris https://github.com/ostris/ai-toolkit
I am not sure which version is best, I wanted to test what longer training did and left it over night. The model is a LoKr.
Description
First version
FAQ
Comments (23)
looks impressive. it seems to break the "professional" photo shoot and turn it more homemade, but still keeping the high quality.
Glad it's working well
By far the best ZImage Lora I used so far. It increases realism without degrading the image quality. Fantastic work
Thanks, feel free to post an example if you tried it. Will be nice to see.
HI! thanx! Is it really LyCORIS? How can you compare the results with lora?
Yes of course it is, when it says it is then it is. Its a LoKr to be specific. I don't understand exactly what the second question is about
@generatorofthingsreal 1) What's better LORA or Lycoris? 2) What your params for Lycoris, training rate, other?, I tried to train it, but with default comfig on 1000 steps seems nothing happens (not training).
@SD_STORM I used a low factor, so a large LoKr model (in terms of size). I trained for 14k steps to see how it worked out, you can see the comparisons in the images. Is it better than a standard lora? I don't know. Settings? Standard AItoolkit
Hi, i'm testing this and it doesn't seem to be making any differences. From what i've read we just use this the same way as a lora right? Or are there any trigger words? I'm using on webui class with zimage
I have no clue about what the webui does these days. Maybe you need to update it or install an extension for lokr to work. Can't say for sure.
@generatorofthingsreal I believe most of the UIs that aren't comfy aren't supporting z-image lokr because it has a unique joined key setup and without explicitly supporting it means only a subset of they weights are being applied.
@ecaj ah, thats that then. Will come soon enough
I've downloaded other files from civit that work with zit, but this one gives me:
[LORA] LoRA mismatch for KModel: /home/xxxx/sd-webui-forge-neo/models/Lora/Z/Large_IF_Zimage_4_000013000.safetensors
using reforge neo. Since it doesn't load, it doesn't make a difference
@boorad yeah most likely a webui/forge issue then. I am sure they will update it soon enough.
@boorad it loads for me, but changing the numbers doesn't change anything. Due note that its a lorica. I did a few readings and it seem to just be used the same as a lora. But user ecaj did give his input.
@generatorofthingsreal is it possible to make this a lora? I've been using the anime2real lora and it works ok. But your lorica gives it a really nice seductive appeal.
@Melodic_Possible_582589 I could retrain it, but that would of course not be exactly the same. Who knew what it would do. Also not something I am gone spend time/cash on rn unfortunately. I am sure forge will catch up soon.
@Melodic_Possible_582589 lokr actually can be better, you may be able to find someone from the community that has patched lokr support for zit, it isn't that hard to do, honestly claude code or others could do it with a little hand holding and pointing to comfy's code.
release full prompts!
check the workflows! but i can post some examples
trained only on women, right?
Over represented at least
At first I had issues with this Lora (and other loras on ZiT), but once you learn how to proper weight (I would not use above 0.7) you get the best results from all the reality/style loras from this one.
Since you use low weights on loras, reinforce what you want on the prompt its a good idea.
Thank you for the Lora, very very nice work.