Anime Femboys for Flux!
Dataset: 110 images 1024xXXXX
Checkpoint used: dev2pro
Trained for 20 epochs, 2 repeats with AdamWScheduleFree optimizer
Why version 0.5? Because currently the LoRA doesn't work that well on its own, without other LoRAs. Yes, it can generate clothed femboys normally, but it fails hard when it comes to anatomy.
I also made the mistake of trusting AI captioning. All image descriptions use "she" pronouns, and words like "woman" or "girl" are frequently encountered. In the new version, which I hope to release next month, I will personally tag my entire expanded dataset.
In this version, I recommend using LoRAs for male anatomy, penises, and all that stuff.
I also used skimmedCFG 0.5/3 or 2/4 or with CFG 2-20. Tested on various gguf quantizations, my graphics card struggles to handle anything above 5_1 or mx6_9 unet and 5_0/fp8 T5.
Trigger words: femboy, chastity cage, flat chastity cage. The last two work inconsistently and it seems like Flux didn't quite grasp their concept.
Possible positives: a digital illustration in an anime style.
Possible negatives: realism, realistic, photorealistic, film, photo, watermark, big breasts, masculine, extra feminine.
This LoRA is my experiment for now, expect an update next month because I ran out of credits on lightning ai. By the way, DM me if you know of any other services where you can train a LoRA for Flux for free, besides Civitai, I have too little Buzz for that, heh.
I don't remember exactly which LoRA I used and its author, any penis lora shoud do.
Ttoo much! TOO MUCH!!!!!!
Description
FAQ
Comments (2)
Try pairing it with simulacrum. It has a ton of native femboy, trap, and otoko no ko in the training set; at least 4500 or more images trained into it with male or femboy. It'll also handle futanari, penis, sex, and thousands of other tags. https://civitai.com/models/803213/consistency-simulacrum-v22-20k-mixed-pack-epoch-5-flux1d-sfwnsfw
It should solidify the anatomy and allow your specific lora to pop. It tends to suffer from lack of focus since I'm bulk feeding it for now, so loras pair well with it.
If anyone is wondering why a new version of the LoRA isn't coming out, let me explain:
1. I've had several training attempts (I train on Lightning AI, they have a certain amount of credits per month), and this (0.5) version turned out to be the best. I don't know, maybe I lack the competence or skills.
2. FLUX is shit. Like, yeah, it's high-quality, photorealistic, makes normal hands and tons of details, but still SHIT. Okay, I'm exaggerating a bit. BUT, Flux is very difficult to train, it requires much more meticulous tagging, and simply - a lot of time and patience... and money. None of which I, a broke student, possess.
3. As of writing this comment, no one has fully managed to unlock NSFW for Flux. And in general, the situation is unstable. Just when you figure out how to use one method, another game-changer is already being released.
4.My 3060ti is not enough for even remotely comfortable use of Flux.
5. Some time after I released this LoRA, Illustrious came out. Words are unnecessary.
To sum up briefly - I've experienced burnout from the inability to properly train a LoRA for Flux.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.