JAV Honda_Riko (本田莉子)版本JAV_honda_riko_v1 (ID: 441096) 综合资源合集综合资源合集

Imitation model. Modelled after Japanese AV star, influencer and youtube celebrity Honda Riko (本田莉子; previously known as Nakazato Sawa 仲里紗羽).

This JAV actress had quite a long filmography, having been in the industry long enough, but had never been ultra-popular. Her first debut was mroe than 10 years ago (2012), similar to the era of Yuna Shiina and Nami Hoshino. She is however still active as of late, but had not been doing many films nowadays. Feature wise, hot body, but her facial look remain quite subjective (and fall short of the recent doll-like Japan AV idol vibes); if you are into slightly more wider lips, and the more 2010's AV vibes, she will be your type. So hope this LoRA does her justice.

~~~~~~~~~~~~~~~~~~~~~~~~~~
Recommended LoRA strength: 0.8 (0.6-1.0 are plausible range dependent on checkpoints; higher values >1 elongates face and widens lips..)

Main Prompt:

hondariko_jav, breasts, short hair, brown hair, brown eyes, bangs, makeup,

variations: single hair bun, high ponytail, medium hair, pink hair

facial variations: parted lips, lipstick, makeup, blush, teeth, smile

*Use of facial prompts depend on checkpoints, her looks signficantly changes with/without makeup. The training images are actually trained with these captions, give it a go.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Checkpoints that work well for this LoRA:

1) Realisian. This is not an easy LoRA to use int he first place, so I think Realisian did relatively well, compared to the rest. Most of the previews are from Realisian.

2) PerfectWorld. Probably due to its 2.5D rendering, PerfectWorld manage to get away with alot of the over-correction that other asian Realistic models suffer. In cowboy shots and anything further than that (full body), perfectWorld does quite well, but the resemblance can be hit-and-miss with closer angles.

3) RealisticMix. Does better with her tied hair/hair bun appearances than the standard bob-cut look. Major downfall is the bad head to body ratio (makes her look like a f**king chibi. LoL)

~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Note for achieving good resemblance:

This model is trickier than what I expected. But let me share why.

If you compare the shape of Riko's face, her facial structure is slightly more rounded than most modern day Asian girl standards of "beauty" (which is somehow a bias incorporated into some checkpoint's training sets). If you run checkpoints like MagiMix Realistic or SHM, you are gonna have a really hard time getting her real face to appear correctly.

Then, because of that, for those of you that had some adequate experience with photography, will understand that people of this facial shape, tend to have a more imbalanced body-to-head ratio (on camera), now couple that with her 155cm height. SInce Gravure photos usually explore these angles, it is not trivial for SD to know how to appropriately correct for those angles in random generations.

TL;DR, there is no quick way to fix this (I tried removing some more unsual dutch-angle shots, but the outcome is pretty much the same), unless you want to spend time adding body LoRAs, negative embedding (which can backfire on resemblance); Or you can just re-seed, which fixes most of these by random anyway~

Now that I have tried to explain how things can go wrong, time to show what you need to do, in few quick steps.

1. Generate the main prompt using a LoRA strength of 0.6-0.8. Now at this stage. you want to pick a weight, that can first mitigate the head-to-body ratio problem. This problem is checkpoint specific, so you will have to try out what works best (hint: lower scale, slightly lesser chance of this being a problem)

2. Now depending on what you use for (1), turn on facial Adetailer, and ramp up the weight there. e.g. if you used 0.6, you can try:

"face of hondariko_jav, red lips, short hair, single hair bun, makeup, lipstick, eyeshadow, parted lips, realistic face, <lora:JAV_honda_riko_v1_epoch_10:1>"

This will get you closer to Riko's look. It is easier to work with the "makeup" version than the no makeup version, in terms of resemblance.

set Adetailer denoise strength to 0.44-0.5. for Full body, you can try to set this higher to 0.55.

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 -0.25 (anything higher will overcorrect face during the process), and turn off Adetailer.

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Issues:

Problem related to understanding modelling bias (explained above), for those that are not keen to know - Bad Image => Perhaps just re-seed the generations.

On Average, "medium hair" works slightly better than "short hair", even though both are just very similar. Bangs can sometimes feel abit unnatural, you can fix that with "realistic hair, fine hair detail"

Post-production edits/Reproducibility:

Some badHand inpaintings.

Happy prompting. And do share [SFW] creations~