Yes, I was planning on doing that at some point.

Done

Same, I wished it was prioritized more. I'm not sure if it's one person or a group but they work on multiple games. This one is always updated the least.

Sure, I trained on 173 images, 25 repeats on every images and 5 epochs in total. I trained it on Anything-v3-1 and anime.vae. When finding photos I try to get as much detail as I can, big resolutions, crops of body, crops of faces, etc. Good pictures are key, if you have bad pictures this will not come out good.

*I train on google colab using the Kohya Trainer. They recently updated and changed a few things and a lot of the wording for stuff isn't the same for the lora gui so i'm going to list stuff off the google colab. Hopefully you can figure out what i'm talking about. I'm also not an expert with this stuff and hopefully someone sees this comment and can help as well. Would love to see better training settings.*

I clean the photos, any png's that have transparency i convert to a random color. Then i use Waifu Diffusion Tagger to make .txt's and then add a trigger word at the end. (You don't need to do this but i see better results).

--Dreambooth Config--
train repeats: 25
reg repeats: 1
instance token: mksks
class token: style

--General Config--
resolution: 512x512
flip aug: false
caption extension: .txt
caption dropout rate: 0
keep tokens: 0
LoRA Config:
network module: networks.lora
network args: n/a
network dim: 128
network alpha: 64
network weight: n/a
--Optimizer Config:--
optimizer type: Lion
optimizer args: n/a
train unet: true
unet lr: 5e-5
train text encoder: true
text encoder lr: 1e-5
lr scheduler: cosine
lr warmup steps: 10
lr scheduler num cycles: 1
lr scheduler power: 1
--Training Config--
lowram: true (probably doesn't matter)
noise offset: 0.0
num epochs: 5
train batch size: 3 (based on pc/gpu)
mixed precision: fp16
save precision: fp16
save n epochs type: save every n epochs
save n epochs type value: 1
save model as: safetensors
max token length: 150
clip skip: 2
gradient checkpointing: false

gradient accumulation steps: 1
seed: -1

These are the settings that I use. Hopefully they work for you and it would be cool if others could also comment what they use. It would help me out as well! (:

@indolentcatmods I use buckets. Sometimes when you crop an image, depending on the way it is compressed, it will lose some of it's quality. Even if you don't see a difference.

@indolentcatmods depends, I usually try to go bigger. It all just depends on the photos you can find. If one photo is bad quality it can mess the entire thing up. Having a small batch of photos is okay and i would start with at least 50 and then go from there to see where you are going wrong. Eventually you can use more photos and see why things are turning out wrong. I also have upscaling off for buckets. I recommend training on squares rather than rectangles. (512x512 instead of 512x768) Also sometimes it can just be the checkpoints/models you are using. Some may not recognize the style well. I train on anything v3 not because it will be the best way to present the style but because other models are based off anything v3. It also is based off the danbooru tagging system so it works well with the .txt's. I found that Corneo's 7th Heaven Mix works well with keeping the house chores style. It's all about experimenting and its not just pressing one button and it comes out.

@indolentcatmods Also you should watch this video about testing. It may be a little confusing if you don't know anything about logic or coding. I put it at the timestamp using x/y generations. It's a really good video to watch in general and an automated way of looking at your lora and seeing what it may generate. Hopefully you can lean some stuff from that.

@indolentcatmods I try to use bigger pictures, if a picture is 1920x1080 I try to use that because its a bigger resolution than something that is 240x240. Bigger resolution means more detail if it is the original image and not a stretched version. You don't need to make these squares just don't train off rectangles. The default should be 512x512 but some people like training off 512x768 assuming its good because that is good for generations. Don't do that because it stretches the images. Hopefully that made sense.

@indolentcatmods I don't crop with a program, I let buckets handle it. When training you have the option to train with different resolution. Keep the resolutions you are training with as a square dimension. Also I don't make closeups myself I usually find them. If I was to browse a website like danbooru for example they have tags for closeup, portrait, cowboy shot, etc. I don't make the closeup's because it would be redundant. The only time you may want to make a closeup is if there is more negative space than positive space. (Positive space is the subject, Negative is the backgrounds and surroundings.) Sorry i was probably unclear.

@IndolentCat sorry for the late response I have been on break. I don't really understand your question. Also i have been having issues with the eyes too. It's a combination of the images being trained and the resolution you trained them on. I found by using the upscale fix (i do it on 2x). It fixes the face generally if you train a model on a lower resolution. It also could just be the art style, for house chores the art style is very easy to copy, but other art styles it may be harder to replicate if there are more details on the face. Again i'm not an expert so im not sure what makes these tick well but i just see that from my experience. I could be wrong so don't quote me.

I'll add it to my queue

i will add it to my list

yes i had that on my list