This model is a LORA for improving humanoid women bodies in SDXL (and by TE in SD1.5)
Usage v0.3:
- main tags are: 1girl, nude, pussy
- important negative: 'huge ass'. If you want natural cat ears, use negative: "headphones, hat" to avoid SDXL create acessories insted of natural ears.
- strength 0.3 ... 0.9 with community models, 1 .. 1.3 with base SDXL. With SD1.5x based checkpoints works only TE (CLIP, lower line in ComfyUI) and good with strength 0.8 .. 1.5
- v0.3 work with MOST of good checkpoints and improves all of what i tested. Best from i tested are: sdxlUnstableDiffusers_v5, v8 (but not v9+), realismEngineSDXL_v20.
- Does NOT work with current DreamShaper XL (alpha nor turbo), sdxlUnstableDiffusers_v9 (and probably all next), very stylized anime models.
- model has over-attention to ass (and some poses), so use it with low strength like (ass:0.5) and/or near last position in prompt. I think this is not my fail: none of photos has big unproportional parts or perspective distortions that can lead to this, so this is probably base SDXL model issue.
Dattaset:
It was trained on ~400 perfect pictures (without any unperfections) with repetition 10 * batch 5 or 2 * 4...8 epoh, and ~500...1500 very good pictures (with most features very good) for diversity with lower attention (repetition 1..4). All womens in range 18-26, most was 20yo.
Trained on base of SDXL1.0fix.
Technical notes:
Best practives and findings:
- Dataset shoud be perfect in all aspects and each feature. Don't think SD has smart AI from what you can prompt good parts and do negative bad, no, but each bad pixel seems do model worse.
- Big diversity lead to problems, i redraw 200+ images by hand to make some consistency of nipples and this improve result a lot but need to improve all.
- Only simple body positions, without many body parts intersections or perspective distortions. No fisheye, macro,
Failed attempts ordered by importance:
- low network alpha lead to very slow training, after 3k-5k steps quality almost no improve, probably need to search best value in range 0.5-0.9 of dim (same as dim mean turn it off).
- scale_weight_norms>3 seems lead to overtrain, 2 is better, 1 is slow, 0 sometimes make interesting results on small dataset but difficult to find good LR and other settings
- Network dim lower then 64 seems not enought to this (or all realistic but not small) dataset
- no-scale of images and train on smaller resolutions buckets often lead to multiplication of body parts in standart LORA (but not tested in Lycoris)
- Continue training lora seems better from "saved state" and with very similar settings. Continue from .safetensor often lead to overtrain some parts. Big changes in settings for continue training often lead model to forget a lot.
- GLORA, OFT seems not as good as they advertised (overtrain with recommended settings). Need more big tests.
- LoKR recommended Multires noise discount 0.8 is not working from start
- regularisation images seems prevent fixing existing SDXL base problems and caused slow training. (need to test for small daaset)
Version history:
all standart LoRA, prodigy preset, scale buckets, no regularization images
v0.3 network 256/128, offset 0.04, 1768 images, fix some images. (tested also 64/64, 129/96, with/w.o./different regularization images, loha, locon and nothing found better, but tests are not full)
v0.25 no spec activation. Partially fixed problematic features. 1750 images, network 128/96, scale_weight_norms=2
v0.2 no spec activation. may or not use 1girl. Works better with sdxlUnstableDiffusers_v5 or v8. 1300 images. learned with prodigy and adafactor preset, network 128/16. Same problematic lips and nipples due dataset diversity and probably overtraining on some pictures.
v0.1 activation: g01tt. Works with ~only sdxlUnstableDiffusers_v5. 1100 images, network 32/16, but has some overtraining issues but undertrained in same time. Problematic nipples.