This week we got 2 slightly athletic ones.
Dont let her baby-face deceive you, on whats underneath that innocent look. Her look is considered average, but its the comfortable kind of average. She puts off a somewhat "amateur" vibe; but her performance is nowhere amateur. Body physique wise, this LoRA will go well with most clothing styles (got lucky that quite a substantial bit of her training data were full body shots - for you stat-hungry JAV fanboys: she's: 93 (G Cup)-61-93cm).
variations: single hair buns, black hair, tied hair
facial variations: natural makeup, grin, teeth, parted lips*
*I do not find the need to use the facial prompts, as most checkpoints recreate her lazy eyes and parted lips perfectly
Checkpoints that work well for this LoRA:
1) Realisian, or any of the Magicmix variants. Should rename it JAV-sian at this point..
2) RealisticMix. Stable, good success; ??????? alittle off, but everything else is fine.
3) ClearRealismMix-Eur, or asian-based/eurasian checkpoints. works 75% of the time with less prompts (i.e. more flexibility).
4) PerfectWorld, or other photoRealistic ?????? models (European model works fine for this LoRA). Tend to over-correct her facial features to be "perfect world ish", but overall still OK.
Note for achieving good resemblance:
Use Adetailer for baseline generation (at 512, 768 etc) only, try to control LoRA strength at 0.9-1.0, adjust your denoise accordingly. For this LoRA, recommended to use Adetailer at ~0.4 denoise strength for close-up, and 0.45 to 0.55 for full body.
For this LoRA, you dont need to set the prompts for Adetailer separately. I prefer to use just the main prompt, and turn ADetailer on, since you dont want to overload the makeup, this actress generally do very light blush and touch-on only.
Recommended approach: is to not use HiRes, but upscale it with img2img instead.
Use a denoise strength of 0.25 (anything higher will overcorrect face during the process), and turn off Adetailer.
Issues:
So far no issues, apart from badHands, and extensive use of inpainting x10 to make the LoRA wear her clothes properly to post this, big 0ppai problems.
Post-production edits/Reproducibility:
Some preview images underwent inpainting manually on the clothing part because of the nature of some of the training data. Prompt responsibly.
Tomorrow, big momma might be ready for release.. lets see.
Happy prompting. And do share [SFW] creations~
v1. Initial release. All epoches perform similar. so the last epoch was picked. retaining last 3 epoch for testing.
trained on 86 images. at 512 reso.
Very good performance on most [asian] checkpoints. rework wont be necessary, I guess.
1. The rights to reposted models belong to original creators.
2. Original creators should contact SeaArt.AI staff through official channels to claim their models. We are committed to protecting every creator's rights. Click to Claim
