Uncut Gems - Z-Image Turbo Edition
This is a moderately successful attempt at creating a uncircumcised flaccid penis lora for z-image turbo.
The model can be used at any strength or cfg, with higher cfg need sometimes to get the shape right, so your outputs may vary depending on the settings you choose.
The turbo model for z-image is not ideal for training new concepts and is difficult to keep flexible while still understanding the trained concept, the base model should allow more flexibility but this model is good enough for now.
The model has been block edited to remove the artifacts but backgrounds can suffer and may need refining after.
Description
This new version responds much better to prompts, things like sitting/standing/hairy/shaved should work now.
Version 1.1 is geared toward natural flaccid and less semi-flaccid like the previous version.
The same strength works for this too, 0.7-1 with 0.8 being the best in my testing.
FAQ
Comments (18)
Pretty amazing model, there were some good models, but it was very large penis and it messes with the eye, thanks for this!
Please make a Lora. The app I used in iPad can handle loha or similar files. And I like your dicks so much 😅
Is it solo male only? on mf couple it gives dick to female too? o.O
For MF it would be best to use a porn merge and then inpaint in img2img
Very good!
Even better is an erection and a small phimosis!
hoping!
Please release as a Lora! Lycoris is not available on DrawThings
👍 I would like to second the request for a LORA version. I am also stuck using an app without LYCORIS support.
DrawThings now supports Lycoris!
excellent work!
Thank you so much for this Lora! Is there any way to stop generating close-ups of the Penis? Almost every image I generate gives me close-ups..
How best to use this?
Use BetterCocks to the normal prompt and at the same time set the Penis Adetailer Model in Adetailer and use Uncut Gems in the prompt within Adetailer?
Would you mind telling me how to add an Adetailer Model in ComfyUI? I haven't found the Adetailer folder.
Yes, at .35-.45% denoise
Could you please make a version for SDXL? A friend is asking for research.
Once the new version is done that's the plan :)
v2 Update:
Not happy with how the test models are behaving, SDXL in particular is reluctant to learn the new concepts without completely overtaking the base model composition.
Going to experiment with adding synthetic images into the dataset, the upscaled images are introducing unwanted artifacts and that seems to be all the lora wants to focus on -_-
Maybe you could ask ppl to donate more photos? Please don’t add synthetic ones in as uncut gems is simply the most realistic one available. I’ve been using it with inpainting and getting way better results compared to anything else I’ve tried.
SDXL version would be awesome but I understand the technical difficulties. How much new material you would need and what it should be like, to get SDXL working better?
Is this model especially incompatible with HomoFidelis 4.0? Using it changes the whole picture. Different background, different person, and the face turns into an ugly mess.