
Imitation model. Modelled after retired Japanese AV star Kaede Niiyama (新山かえで).
Here's another neo-classic that's popular between 2010 -2020. Kaede is another one of those H-cup legend, but her exact filmography is hard to pin-down, having no formal announcement of her retirement (she is mainly active between 2012 - 2016), but did some uncensored AVs with carribeancom slightly later as well. Figure wise, she is considered slightly taller than average (168cm), and fairly muscular. Depending on personal taste, but I would say she leans more towards the "thicc" type, though she do have a baby-face and a tendency to use very little makeup (unlike modern JAV artists).
~~~~~~~~~~~~~~~~~~~~~~~~~~
Recommended LoRA strength: 0.85 (0.7-1.0 are plausible range dependent on checkpoints; higher values >1 lead to distorred facial features, which is not ideal)
Main Prompt:
kaedeniiyama_jav, breasts, long hair, natural makeup, mature woman
variations: tied hair, ponytail
facial variations: mature makeup, parted lips, lipstick, eyeliner, no makeup,
*Use of facial prompts depend on checkpoints, use it if your checkpoint have a bias towards minimal makeup, I actually trained this LoRA with these makeup tags, so feel free to experiment along
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Checkpoints that work well for this LoRA:
1) Realisian. Best in slot - when it comes to consistency, accuracy and expression.
2) RealisticMix has comparable performance, but falls short when it comes to resemblance and posture. Slightly more unstable (seeding) than Realisian.
3) PerfectWorld Is good in a different way, if you are more obsessed with getting the facial expression and face structure right, then go with PerfectWorld. The downside is that you might run into NSFW most of the time, and disproportionate 0ppai. e.g. The preview image where Kaede is in a red Dress, is from perfectWorld.
4) UltimateHentai Realistic or other photoRealistic Hentai models mimics the style of PerfectWorld in getting the face correct, but over-exaggerates her body proportions.
N [do not use]) MagicMix realistic. Too korean-uzzlang like, as the model tend to elongate faces - which I think is the exact opposite of what kaede's feature is like.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Note for achieving good resemblance:
Here's the trick with this model. Kaede's is technically not "beautiful" by SD's standards and the asian-cute-sexy bias inherent in most checkpoint. In fact, what motivated me to model Kaede (apart from mroe big 0ppai JAV babes) is to actually test how flexible are these asian checkpoints in rendering asian faces that is not-the-conventional "pretty". Now that you know the context, here's what you need to do to get Kaede to look as close as what she looks like irl:
1) Remove all "ugly, unattractive, freckles, moldes" prompts from your negative prompt, and sequentially test which negative-embedding is "beautifying faces". Get rid of those, if you have a habit of placing those Neg Embedding.
2) Control your denoising strength for face Adetailer very low. 0.3 is the max, I go with 0.25.
For full body/from behind, depending on angles, try 0.35-0.45. Its very situational, since a high denoising will distort the face. But for most standard portrait, I find 0.3 works well.
3) Slightly reduce the LoRA weight in ADetailer. I find a 0.05 decrease is OK, 0.1 is doable too. This is to control the face slightly, as you need a high LoRA strength in main prompt to model Kaede's body right.
4 [optional]) If your checkpoint do not over-react with weight prompts (like Realisian/realisticMix or any of its predecessor) then add "chubby face" to your main and AD prompt.
An example of the ADetailer prompt I use:
face of kaedeniiyama_jav, eyeliner, mature woman, long hair, <lora:JAV_Kaede_Niiyama_v1_epoch_10:0.8>, realistic face, detailed face, chubby face
PS: realistic face, detailed face is useful to create the dimples, slight wrinkles and some freckles.
Btw, In reality the resemblence is quite decent after you follow these steps, though there's slight over-training, as Kaede's facial resemblence seems slightly better with clothings that are abit more revealing her skin. (cant show it here as they might violate SFW rules)
It is still better to use AD at baseline generation (512,768 etc) and then upscale it later.
Recommended approach: is to not use HiRes, but upscale it with img2img instead.
Use a denoise strength of 0.20 -0.3 (anything higher will overcorrect face during the process), and turn off Adetailer.
I realise for this LoRA, depending on your upscaler, even a denoise strength of 0.2 or 0.3 can make slight difference. test and work what suits best.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Issues:
Nothing that needs immediate fixing, badHands, and occasional bad hand-body ratio. But some of these problems seems to be checkpoint specific.
in img2img upscaling, sometimes the shape of her eyes can be alittle off - I guess you can try running Adetailer again with very low denoise to try and fix that. Or inpaint.
Post-production edits/Reproducibility:
Minor inpainting on some Badhands, and NSFW clothing.
Happy prompting. And do share [SFW] creations~
Might consider training a LoRA just based on her body physique. Lets see how it goes.
描述:
v1. Initial release. Epoch #10. Epoch 6,8 similar, but lets go with 10 for now.
训练词语: kaedeniiyama_jav,breasts,long hair,natural makeup
名称: 417510_training_data.zip
大小 (KB): 55581
类型: Training Data
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
名称: JAV_Kaede_Niiyama_v1.safetensors
大小 (KB): 36989
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success