<h2 id="heading-146">Imitation model. Modelled after retired Japanese AV star Nami Hoshino (星野ナミ).</h2>Here's another classic. It had been more than a decade since her debut (she debuted in 2012) but retired during COVID-19 in 2020. Best known for her oral skills and thick eye shadow, this is one challenging LoRA to model due to her excessive use of eyeliner and eyeshadows. I thought of 'cheating' alittle by trying to train this using semi-realistic (since semiR models tend to hardcode the makeups much better than realistic), but perhaps its my understanding of women makeup thats still lacking. so lets see how this goes.So consider this a beta version, for now.<h3 id="heading-108">~~~~~~~~~~~~~~~~~~~~~~~~~~ Recommended LoRA strength: 0.9 (0.7-1.0 are plausible range dependent on checkpoints; higher values gives a better pre-Adetailer template for the face but can result it very small eyes..)</h3><h3 id="heading-109"></h3><h2 id="heading-110">Main Prompt:</h2><h3 id="heading-147">namihoshino_jav, breasts, brown eyes, jewelry, short hair</h3>variations: black eyes, medium hair, tied hair, earringsfacial variations: eyeshadow, eyeliner, parted lips, lipstick, lip gloss, blush, defined cheekbones*Use of facial prompts depend on checkpoints, use it if your checkpoint have a bias towards minimal makeup, most of the time you only need "defined cheekbones"~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Checkpoints that work well for this LoRA:1) Realisian/RealisticMix/PerfectWorld. A borderline pass. Even so, some generations simply cannot be fixed. But on average, still does better than other models.Perfect world has the slight upper edge when it comes to resemblence, but you get a stronger bleeding with that checkpoint as well. 2) Every other model Similar performance, not very satisfactory tbh, given how well the other LoRA models I have trained had been performing.~~~~~~~~~~~~~~~~~~~~~~~~~~~~Note for achieving good resemblance:This LoRA is one hard model to tackle. I actually trained this model before Reiko Kobayakawa (yup, 3 LoRAs before this), but I can't seem to get it to work to my own satisfactory standards, so I had been trying multiple things (retraining, re-crop, prompting differently), and finally get some idea on how to create her face in a more consistent manner.Okay. Lets break this down further. First, unlike my other JAV models that applies makeup, do not use makeup prompts in ADetailer - use them in the main prompt instead.Thats because Adetailer somehow will try to generate her face without the eyeshadows (even if you prompted it), creating a rather make-up less Nami overall. Second, make sure to control denoising strength for Adetailer below 0.3. 0.22-0.27 is a good range to try out. for full body, maybe 0.35 works - but no guarantee, since the level of resemblence entirely depend on whether the initial generation creates the classic eyeshadow on her face. (if it doesnt there's almost no way you will get a face that fits her usual appearance).third, The solution for this might rely on me spliting this into 2 LoRAs, one for her body, and one for her face. (so far I have managed to avoid doing that, and perhaps this is one good test case, to see if it works). 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 0.25 -0.3 (anything higher will overcorrect face during the process), and turn off Adetailer.~~~~~~~~~~~~~~~~~~~~~~~~~~~~Issues:Currently the facial resemblence is not very satisfactory. Slotting in makeup prompts dont seem to properly address this issue. High potential of retraining.But if you are not so obsessed with creating Nami's makeup look, and have no issue with natural looks, this LoRA is actually not too bad.Ok, I am not convinced, back to the drawing board. I will rework this..!Post-production edits/Reproducibility:Minor inpainting on some Badhands, and clothing.Happy prompting. And do share [SFW] creations~

JAV Nami_Hoshino

v1. Beta version at best. Not stable. But decided to release this and see if there's any kind soul out there happy to thrash this hard for me to gain some useful feedback.

Files with SHA256: e1be345fb10f8bc1a8b905b7e857a3258e183a59dccb9b2e49514f45c76bd844