General futanari penis lora for Ideogram 4
- Works with both natural language prompting and with json format prompting with bboxes
- Can do flaccid and erect from various angles except from behind and male pov
- Works with various bodytypes and ethnicities
- Works with pubic hair when prompted
- Trained on/for futa but works for men too
- Seems to mostly break the "safety" filter
Important: the word "futanari" or "futa" was not used in my captioning, so the model doesn't understand this. Just use "woman" and "penis" in prompts.
Room for improvement:
- Doesn't always strictly follow erect/flaccid when using short prompts
- Does not work properly for male POV (photo taken from man's point of view), it will often invert the penis at this angle. This angle was not included in the training data so this is expected, I may add it when making a new version
- Let me know, feel free to share your results and feedback here
Trained on a dataset of 400 images (60 real photos and 340 synthetic photorealistic images) with natural language captions, not json captions.
Despite not using json bboxes in the captioning, the model seems to "understand" what a penis is with this lora, so you can use bboxes to indicate where the penis should be in the image, you can control size, direction etc with this. I didn't have much time yet to make creative complex json prompts but you can make some pretty interesting stuff with a high degree of control.
For natural language prompting you can use a simple sentence or a detailed prompt, both should work. Prompting in other languages still works, English works best but you can try it out with your own native language.
I'm still experimenting with workflows so I don't have a strong recommendation for lora strength, most of the example images were generated with the lora at strength 1 applied to both models (fp8_scaled), but it can work at slightly lower strengths too. I have not tested how it interacts with other loras and I have not tested how well it works on other quants of the model, only on fp8.
Probably not the final version of this lora, you can follow my account to get notified when I upload a new version.
Leave a like if you enjoy the lora :)
Ko-fi: ko-fi.com/degenerator123
Description
v1 futanari penis erect + flaccid
FAQ
Comments (13)
It's nice to see some new ideogram loras
Zajebiscie!
I'm very grateful to have another IG4 lora, but PLEASE train with bounding boxes (full JSON)! The model was trained exclusively on them, so not using them during lora training is going against the model's instincts. This lora has body horror issues (e.g. double penises), which I'd bet is because you didn't train with bounding boxes.
Also PLEASE don't train using outputs from other models (or CG)! It can make some sense when training a small model with a newer, large model's outputs, but IG4 is open source SOTA, so it's only going to get worse from using outputs from other models. You're baking in their issues and artifacts. In this case, the penises frequently have a strange quality about them, like poor lighting or CG-like shading.
I would also suggest not calling this "futanari" as you say it's not just for women, so it doesn't really make sense for IG4 since you decide who gets the penis with a bounding box.
I look forward to future versions.
1) 99% sure the op used Chroma Flash for synthetic data which is the only existent model capable of doing that, so not lot of choices here
2) Ideogram 4 is not SOTA Open-source because IS NOT Open-source, but open weights. That's also why we got the crappy FP8 weights instead of the full FP32
@Couldvebeenworse Right, open weights not open source. True but most people say open source. Hard habit to break.
@kangaru861 now most people say open restricted weights
Thanks for sharing your feedback and thoughts. I've been making a large dataset of only real images which I'm currently testing. The reason I also used synthetic data for this version is that it is pretty difficult to not bake in the appearance of male bodies, despite captioning everything carefully. Which is also why I include futanari in the lora's title, since it is intended to be able to create women with penises. I only test for futanari so I don't want to advertise it as a general penis lora since I don't know if it works well for men because I don't test that.
@Degenerator123 If you use bounding boxes around the penis when training then it will learn what a penis looks like without baking in the appearance of the rest of the body
please add multiple cock, parallel and sequence
why your lora not breaking with turbo lora? how dd you trained? SNOFS and few other lora break anatomy and style with turbo at 4step 1 cfg but not this lora. tell us your secret
I haven't tested the turbo lora yet so I don't know if this is true. I think the main difference is that this is a rank 32 LoRA, judging by the size I think snofs is a rank 4 LoKr.
an uncut version would be amazing! great work
i' am sorry "Ideogram-4 → Ideogram-Zero: the tool that generates blockages, not images." :)