READ ME, READ ME, READ ME!
How you weight this model is important to achieve max versatility! I HIGHLY recommend you get accustomed to Unet/TE weighting with the syntax <lora:X:Y> Where "X" is the Unet weighting and "Y" is the TE weighting. See example photos for more info. If you don't want to do this you can still get good results, it will just likely take some more generations. Aspect ratio is 512 x 768
This is my third public LoRA, and I put quite a bit of effort into making it. Please reference the example images on how to use the LoRA, and avoid the "burnt" look. It was trained primarily for Otoko No Ko type subjects, but it seems able to handle general futanari and male subjects well, so I've included some in the example images. Anatomy scales quite well with more steps when using Eula A.
General safe weighting: <0.8:0.5> - <0.4:0.4>. You generally want your TE weighting 0.1 to 0.4 below your Unet weighting. I stick to 0.2 - 0.3 below Unet. I also generally recommend using Eula A because SDE Karras seems to get penis anatomy wrong pretty often.
If you want a bulge but no penis peek: stspp, [SUBJECT], bulge, underwear, (erection under clothing), [erection] - " [SUBJECT] " is the kind of character you want: can be otoko_no_ko , futanari, 1boyor something else. ALSO it is often helpful to put penis in the negative prompt if you do not want any penis peek. If you have a male subject it is often helpful to put panties in the negative prompt (unless you want that - I don't judge).
If you want an erection with penis peek: stspp, [SUBJECT], erection under clothes, penis in panties, erection, penis peek, penis, underwear, crossdressing. " [SUBJECT] " is the kind of character you want: can be otoko_no_ko , futanari, 1boyor something else.
Full example prompt: stspp, otoko no ko, bulge, underwear, (erection under clothing), [erection], [school uniform], crossdressing, flat chest, twin braids, solo, sitting, from below, classroom, highres, absurdres, masterpiece, best quality, ultra-detailed, illustration,colorful, masterpiece, best quality, ray tracing, 4k, (extremely detailed), soft lighting, tracing, soft shadows, detailed face, textured skin, <lora:StS-PP-Try4-Long:0.6:0.4>
This LoRA plays well with Age Slider LoRA (for anime), I suggest using that if you'd like to control the visual age of characters generated.
stspp is the activation tag. You can view the TOML I used to train this LoRA here: https://rentry.org/q4uwth . I am new to use the LoRA Easy Trainer scripts, so I am all ears if anyone has advice on how to make this LoRA better :)
Also if anyone has LoRA requests, I'm all ears.
Changelog:
November 18, 2023 - Initial releas
Nov. 18, 2023 - Updated example images to comply with CivitAI's content policy
Nov. 20, 2023 - Updated CivitAI name to be more accurate. Added male example images.
Description
V 1.0
clip_skip = 2
max_data_loader_n_workers = 1
persistent_data_loader_workers = true
max_token_length = 225
prior_loss_weight = 1.0
xformers = true
cache_latents = true
max_train_epochs = 10
no_half_vae = true
`resolution = [ 512, 768,]
batch_size = 1`
FAQ
Comments (3)
I love this Lora! However, most of the dicks are out the top of the shorts. Can you include more data so that a flaccid dick might be poking out from the leg of the shorts? I'll post a couple of samples below. Let me know if this is possible.
I'm currently too busy with IRL stuff right now to do this, but I will add this to my to-do list when I have time. I might also make a PonyXL version of this LoRA while I am at it.
I took a look at implementing this when I update this LoRA for Pony XL, but I could only find about five adequate example images to train off of, so I don't think it's possible to implement this with good results unless you have a dataset of about 15+ quality images you can send me.
I tried with the ~6 images I had and it was pretty hit or miss.