michaelpstanich/

Lora-Breasts_on_Head-big_on_small

Check the "Alpha1 Usage and Information" section for how to use the Lora properly!

WARNING : This lora is far from perfect and requires some work to get creative results. Stable Diffusion struggles with some concepts involving multiple characters but I'll try and explain how to get excellent results below!

Spirit's Lora Training returns to the public space with a pose lora! This is my first ever pose lora and, well, chose quite a doozy. "Breasts on Head" was originally a bounty idea I decided to pick up, though I did deviate from the "Huge on Small with Breast Envy" concept quite a bit. This lora was mostly an exercise to learn more about training poses and how to handle multiple characters. There's still a lot to improve but I want to move onto finishing other concepts and such so I'm releasing what I have as an Alpha. Despite being alpha, it's very and has creative potential!

Just a heads up, since this concept is difficult for SD to understand, the lora is slightly overtrained to make it more consistent. However, the Lora is very easy to break and different checkpoints may require different weights and may require a couple gens to show up properly. I'll cover this later in the description.

< ----- ----- ----- >

Discord - The Broken Chatbox - https://discord.gg/h3vB7S4FEw Donations! =^-^= - https://ko-fi.com/michaelpstanich

< ----- ----- ----- >

Model Name : boobhat_flat_alpha1.safetensors [3ce6f379020b]

Experimental Name : boobhat_flat.ckpt (and -000001 to -000019)

Focus : Assist in creating the "Breasts on Head" or "Boob Hat" pose with a larger breasted woman on a smaller breasted woman.

SFW / NSFW : Training data was mixed with both, but mainly avoided nudity. However, given "breasts” and "boob" are considered erotic the keywords and lora may gravitate toward NSFW depending on the checkpoint. (For example, "SpiritMix - Deep Line" preferred SFW, however "SpiritMix - Rival" tried to pop-out nips whenever possible without prompting.)

Training Contents : Trained on over 480 'anime' images containing a couple different "Breast on Head" concepts. I divided these concepts into boobhat_flat, boobhat_diff, boobhat_close, boobhat_cute, and boobhat_facing depending on the concept. The dataset also has a very heavy bias toward female characters and flat/small breasted women under large/huge breasted women. I excluded images of male subjects (that did not appear feminine) and any furry or furry adjacent images (like cat girls). Bias :

Post Alpha Plans : It may be a while before I return to this one, but I would like to mix up the tag data, re-organize them and make each trained token a bit more specialized while also adding new AI gens to the dataset to help fill some creative gaps. If you'd like to see this sooner rather than later, consider donating to my work and/or leave a comment saying so! ^-^

< ----- ----- ----- >

Alpha1 Usage and Information

Usage Information (Triggers, Tags, Prompting)

WARNING : Given the "flat chest" nature of the concept, the specific poses, and cut-off of most dataset images, checkpoints may interpret the second character as a younger subject, be VERY careful when using this lora in a NSFW manner. (This wasn't intentional and I tried pretty hard to fix it, but it appears that having most of the dataset having one character "below" another has SD interpret the second character as short or very small and I wasn't able to fix this bias)

Triggers

Usage Note : I trained all images onto a single token but divided different concepts onto an "assist" keyword you can use to help steer the output, so you include breats_on_head along with one assist token to get the results you'd want. It, mostly works, though breasts_on_head has a bias toward boobhat_flat because I screwed up the repeats so it's over-represented.

breasts_on_head Always include this trigger when using the Lora for the best results. In alpha1 the entire dataset was trained with this keyword included and has the most weight, however it also has the heaviest bias so you can reduce the weight of this trigger word to potentially improve or change results when increasing the lora's weight or weight of other triggers.

boobhat_flat Heaviest concept in the dataset, use this when you want a small/flat chested girl under larger breasts. boobhat_flat can also be introduced to the negative to help un-bias the breasts_on_head trigger but results will vary based on checkpoint. Was trained on images where the second character did not have larger breasts, or they were covered and implied a small size.

boobhat_diff Use when you're not specifically targeting flat/small breasted for the second character, or would like the 'above' woman to not have huge breasts, for example, one large breasted and one medium breasted. This was trained on general girl/girl breast on head images where breasts sizes are descernable and has the most pose variety. (Can also be useful if trying to to a creative pose with a flat chested girl)

boobhat_cute Use this when you don't want breasts on a character but instead a plushy, stuffed animal, stylized chibi, or another object. This trigger is rather inconsistent and requires some additional prompting to get working on different checkpoints. Was trained on images where the breasts rested on a chibi, very small child, or small creature/object.

boobhat_facing Use this if you don't want the characters facing the same direction, can combine with other assist triggers. Was trained on images where the characters were either hugging or looking in different directions, is fairly inconsistent and may require increasing the lora's overall weight. (This aspect of the dataset needs more resources)

boobhat_close Trained on images where only the below character's face was visible. Not really super useful and was only trained to prevent bias' in breasts_on_head but may be removed in future versions. (maybe people can come up with creative uses?)

Important Assist Tokens (Add these to help get the concept to show up more consistently) Positives : 2girls, woman, girl

Optional Assist Tokens (Use based on what would help your concept be understood) Positives : breast envy, looking up, nipples, hug, hug from behind, hugging, breast rest, squish Negatives : nipples, hat, hug, hugging, grabbing, touching, 1girl, solo, yuri

Usage Tips (General) Since I made this to be more flexible than a single static pose prompting is a bit more complex and not always consistent, so here's some general tips. When prompting try to always include tags for what you want to see, such as "2girls" or "twins" when doing multiple characters or "1girl" and "solo" when using boobhat_cute for objects. When prompting in the characters the tags were lightly orders with "woman" followed by the above girl's related tags like her size and clothing, then followed with "girl" and the below girl's related tags, then followed by environment. So for example, if I want a huge breasted woman in a t-shirt and jacket on an annoyed medium chested girl wearing a dress I would prompt something like... " boobhat_diff, breasts_on_head, woman, smirk, huge breasts, green eyes, long hair, shirt, jacket, girl, annoyed, short hair, medium breasts, dress, indoors " Note that the lora's trigger words for Alpha1 could be re-interpreted by a checkpoint as well, so you may need to prompt against a checkpoint's own bias for breasts on head, as well as maybe include "hat" in negatives if it keeps giving your characters head gear. If you start including different view directions or poses the lora can break rather quickly, you can increase the lora's weight as well as the trigger keywords to help offset this.

Advanced Tips (controlNet and advanced pose mixing) But let's say your pose isn't working, this is where ControlNet comes in. Without the lora some models can actually create this concept on their own as "breast on head" is actually a well known tag, however it's often rather polluted and using the lora allows any anime checkpoint to re-create the pose. Here I'll explain a process to get this rolling. First you may just want to roll with RNG, try to generate images in a certain pose while switching around tokens, if you get something close or creative you'll be able to put that image in controlNet. I used a mixture of openpose and softedge with lower weight and early ending times (usually between 0.3 and 0.7 weight, and ending at 0.15 - 0.3 on a 50+50 hires step process). ControlNet will now start the gen with that pose and the Lora will then fill in the gaps. You can then use this to change view angles like changing from a low-angle shot to a "from above" shot. (I'll post a graphic to display this) This method uses the Lora as an assist lora rather than for direct generations, and you can always use this method with normal images you found such as maybe there's a breast on head image you found somewhere you would like to replicate but normal controlNet struggles.

< ----- ----- ----- >

What this lora does well :

What this lora struggles with :

Other Quirks / Issues :

< ----- ----- ----- >

Hey there, so like, this stuff takes a lot of work and I have many other projects, if you'd like to support my work then why not consider a donation? You can send me donations over on Ko-Fi! https://ko-fi.com/michaelpstanich

If you have any suggestions, feedback, or ideas you want to throw my way, head over to our community Discord, The Broken Chatbox! https://discord.gg/h3vB7S4FEw