Usagi Tsukino (Fanart) - Sailor Moon (LOCON version)

Note: This was trained mostly on Kuroinuu Juu and Glamour works so it has a bias towards bigger breasts. Don't complain if you try to make her flat and it doesn't work.

V2:

Here's the V2 as promised, a bit late because there was a power outage minutes before saving the last epoch. Was the universe against it? Likely! But as a deviant proud man of culture, the universe can go fuck self pleasure itself :P.

Works fine with: lora:SMUsagiTsukinoV16:0.65 smusagitsukino,

I managed to more or less isolate the following outfits, I had to be stupidly specific to maximize separation as i kept getting a virtual soup:

• ClassicSM_Blue_skirt_White_gloves_Red_boots_Red_choker_Red_bowtie_Red_back_bow_Circlet_Red_bun_cover

• EternalSM_Yellow_layered_skirt_Pink_puffy_sleeves_White_gloves_White_boots_Wing_brooch_Red_choker_Red_ribbon_Wings_Hairpin_Red_bun_cover_Golden_crescent_forehead_mark

• SuperSM_White_skirt_White_gloves_Red_boots_Yellow_choker_Red_bowtie_White_back_bow_Crescent_circlet_Hairpin_Red_bun_cover

• Princess_white_strapless_dress_huge_back_bow_Golden_crescent_forehead_mark

• school_uniform_blue_serafuku_blue_back_bow_red_bowtie

• golden_crescent_forehead_mark: use this in the negatives to get rid of it or boost it.

This V2 is technically V16 while the prodigy optimizer is not normally better than AdamW. In some edge case like this where the sailor moon concept is so cooked into the model it overbakes anything even tangentially related to it.

So here's what happened, the eternal outfit is much more complex that the other ones yet is still recognized by SD as a sailor senshi outfit. So to train it properly I need to increase the repetitions but the sailor moon concept is so trained into the model that the extra repetitions overbaked the LORA. The V1 iteration of the LOCON was done using normalization to limit the over bakedness but it only worked to a point. I spent most of this and the past week tweaking the repetitions, tagging and leaning rate, until i stumbled on the Prodigy optimizer. As it is dynamic it seems to have done a good job at leaning the undertrained parts(the outfits) while ignoring the over trained parts(usagi herself.)

Regardless! As always I'll add the training dataset and tags summary so people can play with those. There's a lot of bits and pieces of her uniform in there if you want to mix and match instead of only using her stock uniform.

V1:

Don't you have one of those occasions in which you know the answer to a problem, you actually have two plausible solutions and yet you stupidly take the third, the long shot one?

Well this is it. This LOCON uses mostly the same dataset as my LORA version with some minor tweaking.

Seems to work fine with: lyco:SMUsagiTsukinoV6:.6, SMUsagiTsukino,

Available clothing sets are:

• StandardSenshiOutfit

• EternalSMOutfit (Bleeds over super)

• SuperSMOutfit (Bleeds over Eternal)

• SMPrincessOutfit

• SMSchoolUniform

• KuroInuLingerieSailorOutfit (porny)

Note: This is not better than the LORA, it is just different. Only "positive" thing I noticed is that If you use it with a photorealistic model, it makes dollar store cosplayers.

The issue with my LORA was bleeding between costumes, it is till present here but in a lesser degree. My plausible solutions were either isolate the costumes as if the were different characters essentially making the LORA a character pack or B use regularization images to make sure the LORA understood that Usagi Tsukino isn't the same as sailor moon. Well I choose the third option, my thought process went like this: "LOCON absorb more detail, thus even if LORA sees them as the same maybe the LOCON won't" and it kind of worked. Her default outfit and the Kuroinu's porny one have a lot less bleeding. Super and eternal are sill a soup. I blame the skirts multi color nature making them absurdly similar tag wise in SD's eyes making them essentially the same except for some coloring and small accessories.

Anyway... at some point I'll do a retrain, probably isolating the outfits as individual characters, if I have the motivation i might even do a V2 for both the LORA and the LOCON. I could try reg images but the extra training time coupled with my inexperience makes me reluctant. The theory seems sound in my mind, in fact i already had a small group of images tagged only as Usagi tsukino and expected SD to sort the tags by itself, but they were a relatively small amount of images compared to the full weight of the others, if I use them as Reg images I'll have to up the repetitions by a lot. and probably double the already long training time. But I am rambling let's leave it as regularization images seem like a solution but they always bite at your available training time.

As always I'll add the training dataset and tags summary so people can play with those.