It have to draw textile-free (nude) female humanoids.

Usage:

1. connect node with this lora after checkpoint v5-“SDXL Unstable Diffusers ☛ YamerMIX” https://civitai.com/models/84040/sdxl-unstable-diffusers-yamermix

2. set strength of model and clip to 0.6-0.8 (sometimes better so adjust separately to fit composition)

3. use (or none of them) activation words to attract women with different faces:

   • g01tt, = general

   • g3475, g151a, g159a, = some characters, but now model seems undertrained and it not affect very

4. use negative to avoid mutations and style picture (see example picture):

• detailed:

out of frame, lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, duplicate, wrong proportions, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature, easynegative, bad hands, (low quality, worst quality:1.3), backlit, bad-picture-chill-75v, corrupted image, letters, words, censored, multiple fingers, multiple legs, bad hands, blur, pixelated, bad anatomy, cropped hands, artist name, extra digit, loli, fewer digit, author name, names, autograph, hand writing, written text, five legs, five legs, 3 legs, three legs, five legs, 5 legs, bad hands, bad legs, bad body shape, multiple tails, two tails, deformated body, imperfect painting, blurry hands, blurry face, blurry body, blurry image, bad eyes, ugly face, ugly eyes, repetition, ugly body

and

cartoon, comics, anime,

• vivid:

(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation. tattoo, grayscale, frame, border

and

cartoon, comics, anime,

• general:

ugly, disfigured, deformed, bad anatomy, comics, cartoon, anime, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, grainy, signature, draft

some poses need to negative

(extra legs, vaginal penetration, penis, sex,)

as dataset contain few (~5%)

Early results:

• characters do not shy to show their nudity and it not bad-looking in ~25% of renders

• you can gently ask her to spread legs or even insert something

• she understand few body poses as “SDXL Unstable Diffusers ☛ YamerMIX” do.

• i feel it can produce good pictures if tweak it more, but now img2img with best 1.5 models will improve quality and do stable good results.

• starting with strength 0.6 this Lora attracts mostly nude girls to pose in composition

Cons:

• due to base model (YamerMIX v5) it has over-expressed smiles And tends to anime/CG style, that hard to normalize with negative (comics, cartoon, anime) .

• it produce many images with wrong legs quantity, useless without proper negative prompt.

• steps under 40 produce weird artifacts (see example pic)

• it is much worse then best 1.5 models, but it can produce beauty images.

• often continue mutant-body under buttocks, probably due (untagged?) mirrors and water reflections in dataset

• it almost always contains some text…

• it uglyfy composition a bit, it can uglify characters

• it can’t understand difficult poses (legs above head, crossing legs, arms on cheeks, itc)

Tech:

Trained on 1100 very-best photos that i collect whole life with different train repetitions according to my aesthetics sorting and to avoid big photo-sets trained more that only 1-image persons.

Trained on top of base sdXL_v10VAEFix, but better to use on “SDXL Unstable Diffusers ☛ YamerMIX”. I tried train based on other models and results are a worse, need more tests.

it is undertrained, or dataset it soo wide or lora size is small, or whatever. Anyway next 5000 steps make it worse. This was Prodigy preset, but tried all presets in Kohya, probably need to lower learning rate and test again. Tried dataset with 100 very-very-very best images but it became worse after 750 steps in every preset, AND same time it does not learn well key tags till 1000 steps.

Probably upscale buckets is not good – need to investigate.

TODO:

• try to train SDXL refiner

• try train again with slower LR and cos

• clean dataset from text (now i spend 1-3 minute per image) + delete text-related tags

• upscale low res images (now 10% lower then 768) using SD1.5 and make better compositions and pose diversity

• better normalize dataset to avoid tends to single face

• train big model (checkpoint). Seems like need more then 24gb 🙁 So need to find a job and work a few years…

If you want to help in some way – welcome. I don’t have discord, tg, or other acc.