Trigger word
None required (fully trigger-free) The LoRA activates purely from natural language descriptions.
Please see my example photos for an idea of what is working currently.
It seems to default to more Asian looking women unless you specify a look.
I'm working on a larger dataset to increase options as it feels limited in some areas but when you do get the results they are amazing! I love this new model and how to treats facial expressions compared to PonyXL ect.
Base model
Z-Image Turbo (all samples generated on pure Z-Image Turbo)
Generation info for all sample images
Z-Image Turbo, Euler, 9 steps, CFG 1.0, LoRA strength 0.4–1.2
UPDATE: 0.4 weight seems to be better with keeping face information and mixing with other loras.
What this LoRA does
Trained on a hand picked dataset of the most photorealistic tentacle scenes I could find. Extreme cumflation, deep oral/anal/vaginal, tight wrapping, suction cups, ridges, slime, post-climax overflow).
Recommended Settings
Base: Z-Image Turbo Sampler: Euler Steps: 9 CFG: 1.0 LoRA strength: 0.4 – 1.2 , lower the weight if using with a character lora.
☕ Support
Enjoying my Loras? or want custom commissions? [Buy me a coffee]
Help fund more training using my runpod referral link, We both get credit! [Runpod ]
Description
FAQ
Comments (11)
I'm continually astounded by how quickly Z-Image picks up on complex subjects. A v0.8b version, trained so early after the model's release, shouldn't make this much visual sense. It just takes so long for larger models to pick up this much accuracy.
I know, Its so easy to work with! I tried doing similar things when Flux1 was released and it was next to impossible. The fact that all this is being trained in only a few hours on a consumer GPU is crazy!
@BalakayMotion How are you training it?
@BalakayMotion That is crazy. Z-Image is getting an absolute avalanche of loras as a result, and they're generally good. I think it's clear that we have a new SDXL, after I thought that we could never get that kind of adoption again. I assume you're using AI Toolkit? Do you have any tips on training? Did you use detailed captions? What learning rate? Resolution?
@Jellai Yea I'm using ai toolkit for this one its Rank 16, 6000 steps. LR 0.00005 for 35 image dataset! Lots of room for improvement! I'm going for another run with a larger dataset.
Can't wait until we start getting some checkpoint merges like CyberRealistic that will hopefully bake in some NSFW fixes ect.
Thank you author, I tried LORA and it worked, but I found that it modifies the face, for example, I drew a picture of XX celebrity in XXX, it directly adjusted the facial features of the person, resulting in no longer resembling XX celebrity (if you look closely, you can indeed see that it is similar to the SDXL when it finally generates a star picture to fine-tune the facial features to prevent XX celebrity from being drawn directly). What is the reason for this? Is it because the main model I use is Q8.GGUF? Does this brother LORA have to be combined with the BF16 E4M3 quantization model to achieve the best effect?
I'm not 100% sure, I have some personal trained celeb LORAs that are working fine however I do have to prompt ethnicity as this is currently defaulting to a more Asian look. Some images in the dataset need altering for my second run I'm planning to do tomorrow so look out for the next version. All my examples are using the BF16 model so that could be effecting your results but doubt it.
Ok, I have found this model works a lot better at lowe weight 0.4 even,
@BalakayMotion You have to use your character Lora loaded with LoraLoaderModelOnly (without clip) and use the native Load Lora node (with clip) for the tentacles. The sweet point is 0.5 - 0.6 strength_mode and 1.0 clip. Use Euler A & beta for Z-image for fast renders. Any DPM is expensive in resource and time consuming.
@sczakana THX~
Looks great, going to give it a try. Thank you!