For the best experience, use it together with a futanari friendly checkpoint like Futanari Factor
This is an "improved" version of the FutaVeinyV6 that has more penis and testicles sizes. The adapter was trained using LoKr and was tested together with models that are already able to draw futanaris without much effort like abyssFuta, bb95-furry-mix, fluffyrock and their merges.
How to use
Edit the "SIZE" to the size you want to use.
Penis sizes:
small penis (WIP)
penis
large penis
huge penis
gigantic penis
Testicles sizes:
testicles
large testicles
huge testicles
gigantic testicles (WIP)
If the size slider doesn't seem to be working well with your prompt, try adding some weight to the token:
futanari, veiny penis, (gigantic penis:1.4), (huge testicles:1.2)Optional:
You can add "erection" or "flaccid" to your positive or negative prompts to control the dong status.
(WIP) If you like some saggers, add "sagging testicles" to your positive prompt.
(WIP) Add "foreskin, phimosis" to your positive or negative prompts to change the dong sheath.
Your feedback is appreciated! Its going to help a lot with tracking what checkpoints are working the best to produce the best dongs and what I need to change in the next iterations of this lora.
Development notes:
The training images were tagged using conv2 and swinv2 at the same time and then merged together and the duplicate tags removed.
Every caption was revised by me and the concept tokens were reordered to be used with keep_n_tokens.
Every image was cleaned of signatures and watermarks using lama cleaner and photoshop.
Censored images were redraw using deepcreampy and then manually retouched using manga studio.
5%~10% of the dataset is reserved for AI generated images made by me that were upscaled and manually retouched using manga studio and a drawing tablet. It helps to augment some concepts that don't have enough good pictures and achieve that specific futanari aesthetic I'm looking for.
I took 1 month to assemble, clean, redraw and tag the dataset, but the training and testing process is even more time consuming. To help with that, I'm also trying to code a tool to help me evaluate the testing grids, find the best epochs and track what worked or not between batches.
Description
Improved compatibility and overall aesthetic.
Trained with dylora so it could be resized by someone willing to experiment on it
FAQ
Comments (12)
could we get a reverse concept - so a trans man? no boobs, no penis. I cant find any whereas theres a loooot of futanari loras lol
I don't think you should ask someone who likes the opposite of what you described. Also futa and trans are not connected and have different origins.
There was a lora called "cunt boy" for quite a while floating around. It was trained exactly on the subject you wanted, but it was made using clip skip 2 and NAI, so it was only good for 2d pictures. I couldn't find it anymore in here, but with some luck I think you can still find it with some google wizardly
Can we have a different size options? like small, tiny, micro...etc
I'm still working in curating the dataset for each size. Right now it's really unbalanced, but it should do better in the next iterations. It's going to take some time since I'm doing it manually, inspecting picture by picture and tagging sizes manually.
Every time I try to run this I get "RuntimeError: Given groups=1, weight of size [1, 2880, 1, 1], expected input[1, 320, 96, 80] to have 2880 channels, but got 320 channels instead"
How up-to date are your web-ui right now? It seems like compatibility issue with the lycoris module this lora was trained on the client's inference lib
@ckmlai Not very. I guess that'd probably be why.
could be a great Lora/Lyco. However it basically always generates this ugly and weird glans/tips.
If you can get that improved....
Sorry! This one is supposed to be used together with this checkpoint here https://civitai.com/models/110541/futanari-factor or at least some checkpoints with prior knowledge of futanari concept like furry stuff
@ckmlai Yeah, I actually did use it with Futanari Factor, and 2, 3 other Futa-based Checkpoints. But still...
@daMatt By looking at your sample, I can see that you're using natural language and some realistic embeddings, both of which don't mix well with loras and checkpoints trained on booru tags and clip skip 2. You can check the embeddings, checkpoints and tokens I'm using to generate my pictures by looking at the generation parameters on civit or by just drag and drop one of them into your web-ui. You can use those to compare and see what I'm doing differently and modify your prompts accordingly