Type | |
Stats | 413 |
Reviews | (45) |
Published | May 8, 2024 |
Base Model | |
Training | Steps: 3,204 Epochs: 12 |
Usage Tips | Clip Skip: 1 Strength: 0.9 |
Trigger Words | ayumishinoda_jav breasts long hair wavy hair parted lips |
Hash | AutoV2 D31F73609C |
Imitation model. Modelled after retired Japanese AV actress Ayumi Shinoda (篠田あゆみ).
Many of you probably dont know this one. So here's a brief intro - Ayumi gained alot of popularity between 2014 - 2016 (her time in the industry is really short), from which her filmography is insane. She announced her departure in Twitter mid 2016, so she did not really have a "graduation film",unlike many other JAV actresses. She previously debuted under Miwako Ikeda, but film volume wise, she did more under the alias of Ayumi. She did quite a fair bit of fendom, BDSM, MILF, hardcore, cosplay fetish with a very developed and mature body despite not ranked very favourably in the looks department. She also won the Bets Mature Actress JAV award by SKY PerfectTV (prior to her 30th Birthday) in 2015.
I decided to work on training her LoRA, after the whole episode with Nami Hoshino, where SD seems to struggle with "japanese" (squinty/down-turned eyes). So I tried to minimise the makeup tokens in the training data - And this time the results look much better.
~~~~~~~~~~~~~~~~~~~~~~~~~~
Recommended LoRA strength: 0.9 (0.7-1.1 are plausible range dependent on checkpoints; higher values >1 creates the downturned eyes but also increases aging..)
Main Prompt:
ayumishinoda_jav, breasts, long hair, parted lips, lipstick
variations: single hair bun, ponytail, tied hair, brown hair, wavy hair, glasses
facial variations: parted lips, lipstick, makeup, blush, teeth, smile, partially closed eyes
*Use of facial prompts depend on checkpoints, I didnt actually bother, since the main prompts works fine. (see tips below)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Checkpoints that work well for this LoRA:
1) Realisian. Best render so far.
2) Perfect World. Similar performance to Realisian, slightly more washed up colours. but good degree of resemblance.
3) RealisticMix. Too NSFW. abit hard to manage bust size.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Note for achieving good resemblance:
The trick here is to generate
down-turned eyes are by default labelled as a somehow unfavourable feature in native SD, and many other checkpoints. I tried using negative prompts, doesnt quite work, as the model will still try to correct "ugliness".
The solution turns out to be quite easy, force a generation with the main prompts (using very high weights on the LoRA ~0.9 -1.0), then turn on ADetailer, but set the strength to only 0.25 max.
This will work most of the time, unless you are doing a full-body shot; if thats the case, then run ADetailer with a separate prompt, and set the strength slightly higher to 0.35.
If you over-correct, the down-turned eyes will become generic "beautiful eyes".
Every checkpoint is slightly different, so experiment incrementally.
Very straightforward to use. If you are lazy, you can just use the main prompt to control for everything and apply Adetailer to fix the faces.
Else, you can separate the prompts into 2 stages.
1. Generate the main prompt using a LoRA strength of 0.8-1.0. so its probably safe to say that this weight is stable for use.
2. Now depending on what you use for (1), turn on facial Adetailer, and put something like
"face of ayumishinoda_jav, long hair, parted lips, teeth, <lora here>"
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. I prefer 0.2 since I use foolhardy-reMarci.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Issues:
No issues so far, apart from some badHands (in some checkpoints).
Post-production edits/Reproducibility:
No edits. all img2img original upscales.
Happy prompting. And do share [SFW] creations~