Someone on Pixiv requested me to make some illustrations of Juri, so I put together a LoRA for her.
Trained on 37 images, 10 repeats. 16th epoch was pretty good but it missed some of her hair and dress features, so I picked 18th epoch as the optimum. If there are errors in generation you can try to lower the weight to 0.8
The examples I show all use a weight of 1.0
Description
Trained on NAI
FAQ
Comments (20)
I posted also one for Makino but it's under review. Do you have a Discord?
I have a discord but I almost never log in. What do you use discord for?
Civitai is being weird, I can see your Ruki post on your profile but when I click on it, it says I can't see the post because of my filter settings (and all my filters are off so idk why)...
@FaeFlan Dm mainly, but we can discuss here if you prefer. Idk if my Makino is visible now...
@Gaspatcher Nope, still can't see the post when I click on it. My discord is flutter.mane (just logged in today and set a new user name)
@FaeFlan I've just added you.
@Gaspatcher Yup, got your request :) Added back
Could you try adding D-Reaper Juri into this LoRA, or would it be easier for you to just train a new one? Been trying to train my own but haven't been getting good results.
It's probably easier to train a new one. But there are basically no fan art of D-Reaper Juri so someone has to take enough screenshots of the anime to make a good dataset. Unfortunately I don't have time to browse through the episodes where she appears...
Do you already have the dataset to train with? What problem are you having with the training exactly?
I have a dataset ready, but it's very small, 15 images for her human-like form and 10 for the one with blue skin. Like you said, there isn't a lot of fanart, so I added some screenshots from the anime and some screenshots from Koikatsu so I could have more images to train on. This does add the risk of the LoRA's artstyle being like the anime or Koikatsu, I've been seeing it a bit in my epochs. I can send you the dataset, and if you need I can take more pictures of the anime or Koikatsu.
Main problem I've been having for training is that I only have a laptop with 8GB VRAM, so I can't change settings a lot without running out of VRAM. Also most of my training has been either over or under-trained due to how small the dataset is. I'm wondering if putting in pictures of normal Juri would help or not...
@c18h27no3 Yeah, the dataset is probably too small. I wouldn't mix the normal Juri with the D-Reaper because not just the skin is different but also the eyes if I recall correctly. I would get at least 30 images of the D-Reaper Juri from the anime if I were you. Different angles and different backgrounds.
In theory you can input your successful generations into the dataset, I have been doing that for some of my Neopets LoRA since those have very little images and fanart as well. But usually it takes a lot of manual cleaning because the generated images typically have really bad errors included.
After looking through the anime again I couldn't find nearly enough screenshots. I added 11 images, a lot of which are close-ups and back shots. Also didn't find any old gens that were good enough to justify putting them into the dataset. I'll try training and see if that makes a difference, would you be willing to take a shot at it if I send you the dataset?
@c18h27no3 Sure, I'll give it a try some time this week. How will you share the dataset?
The size is a bit under 8MB, do you prefer Discord, Google Drive, or Mega? Was also going to suggest Catbox but they seem to be down right now.
@c18h27no3 Does discord compress the images? If it doesn't you can send me it through discord, my username is flutter.mane
If it does compress, google drive is fine too
I've sent you a friend request on Discord, my username is c18h27no3#7844
may I ask if there have been any updates on this?
@LolLord You mean D-Reaper Juri LoRA? I posted it already, you can find it in my profile
@FaeFlan oh thanks, I had an open search for "digimon" and havent seen it there so I thought it didnt exist yet, but sure I guess it doesnt have digimon written anywhere :-)
Leomon sama XDDD!