Would appreciate images, feedback, requests, reviews, and of course comments in the comments. Please check out my other models if you like this one.
LINK TO MODEL AT BOTTOM
I've noticed the eyes on a few of my Loras don't look great, if anybody knows how to make them look better I would appreciate it.
A Lora of Suzuya from Kancolle.
had a single image with weapons which decided to become a nuisance and pops up quite a bit in generation (will retrain without it).
accidently baked in white_shirt and red_bow, but doesn't seem to effect it that much. (if it is affecting it try putting the tags in the negative)
mixing them gives quite good looking results (check final five examples of either version - gave me an error when trying to upload the final image to AOM2 version)
Link to model: MightyCorneo (new merges with MeinaHentai) | Stable Diffusion Checkpoint | Civitai
Description
trained the same way but with NAI
FAQ
Comments (5)
For the eyes issue do you have any examples? Just offhand, adding finely detailed eyes/iris/pupil seems to help me but it still relies on the model knowing what to do. So it could be a prompt-fixable issue or it could be a problem with underlying datasets.
I think it's probably a dataset issue; do you include close ups of the eyes in the training, because I haven't been doing that much and I think that is probably why.
@bloodsplash09 it may help a bit, but if you can not find a consistent artist which focus on face, a closer training data may be a drawback. From the information so far, it maybe better train the facial data after you finish the base checkpoint.
Also, use more trigger tags on dataset (e.g. instead of using character name only, use prompt to describe the character) will help to increase the detail, but will make it difficult to trigger up. Network Dim size also helps a bit, but it depends on your dataset.
@4BEraser thanks for the info, making it more difficult to trigger is a slight drawback but it's better than a half-decent Lora. I also noticed that sometimes the training won't bake in the outfit, even if it is the main bulk of the dataset and not captioned (probably just an error on my part). It's sucks not having extremely simple step by step processes for a good Lora. I've seen many people saying "do this, do that" and I don't really know the parameters that well, I'm just using a config I've stuck with for a while. Spent around 7 hours repeatedly trying to train a lora and some of them were better and some were worse, but I didn't like any of them. Got a bit demotivated as no matter what I did I could only get it nearly where I wanted it to be.
Sorry for the rambling, I'm tired and annoyed lol.
@bloodsplash09 yeh, troubleshooting is always a mess, and I have to write a long introduction for everyone to use the long prompt easily, otherwise it will fcked up for no reason.
Still, the plus side of more tags is that you barely need to adjust any tag of it, just let it go all random. The time will mostly cost on solving a best dataset. From my experience, over 50 pic of 1 concept is enough, but more high quality pics, with different angles are always better. For a chara concept has 1000+ pics, I will pick about 120-150 of them.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.