<h2 id="heading-146">Imitation model. Modelled after Japanese AV star Monami Takarada (宝田もなみ aka I-cup goddess of JAV).</h2>The actress is not best known for her looks (then again looks, is not necessarily the only thing in the JAV industry ~.~). Trying to train her facial features, am pretty sure, had been a challenge for SD (and me). Still, I need to make her I-cup justice. Many epochs overfitted her features, and training-loss stats were also misleading; so had to manually test everything.But, I think I got it as close as I can. <h3 id="heading-108"> Recommended LoRA strength: 1 (0.7-1.1 are plausible range dependent on checkpoints; reserve lower values only if you are struggling with other LoRAs)</h3><h3 id="heading-109"></h3><h2 id="heading-110">Main Prompt:</h2><h3 id="heading-147">monamitakarada_v2_jav, large breasts, brown eyes</h3>makeup prompts to use: heavy makeup, red lipstickvariations: bangs, smile, parted lips, brown hair,Checkpoints that work well for this LoRA:1) ClearRealismMix, this model is the clear winner, I am not sure why. But it just does not attempt to over-correct her facial feature, which is key for this actress.2) Realisian, does pretty well, with the caveat of smoothing out her face (which is not too bad, considering this actress sometimes post heavily edited selfies @.@ )3) RealisticMix, or other photoRealistic asian-eur models that respects facial prompts well.<h3 id="heading-1289">Note for achieving good resemblance:</h3>Context (skip for TLDR): This actress has a fairly unique facial feature. First, she dont follow the usual slim-slender facial structure; she leans more towards the chubby side like (Alice Kisaki). Yet, her eyes are slightly "squintier" than average (I tried using prompts like "asian eyes", but it just made it smaller not squinter, until I retrained her.) In this released version, you dont need to do anything, just set LoRA strength at 1, or if applying loRA set the LoRA strength to 1 at the face Adetailer prompt instead.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 ~0.45 to 0.55 denoise strength for close-up, and 0.45 to 0.60 for full body.You will also need makeup prompts, if you do use them, you only need the following in face Adetailer:face of monamitakarada_v2_jav, lipstick, &lt;lora:JAV_monami_takarada_v2-000006:1&gt;<h3 id="heading-1290">Recommended approach: is to not use HiRes, but upscale it with img2img instead. Use a denoise strength of 0.25-0.35, and turn off Adetailer. I find this is one of the ONLY WAY to retain her unique facial feature.</h3><h3 id="heading-1291"></h3>Issues:Consistently across most epochs, there's widespread issues with generating "good hands". There is nothing unusual in the training data, so I am still not sure whats causing this. One way to address this is to use Good Hand LoRAs, seems to mitigate bad hand generation though not perfect. Either re-seed for a better generation, or artificially tweak this with Adetailer post-production.Happy prompting. And do share creations~

JAV monami_takarada

v2. Initial release. at Epoch #6. Version that is trained using WD14. 92 images (v1 using BLIP is obsolete). Idea is to overcook the training, then work backwards.

Files with SHA256: 1232bfd6fab76e833a2876f6e3ea1229d268d52e12b8087e28845a5bf43de896