Recommeded prompt: ht_renge, adding medium breasts for her aged up version
Outfits:
Bikini:
black bikini, side-tie bikini bottom, hair tubes, white hairbandBlack dress:
black dress, choker, short sleeves, goth fashionKimono:
black kimono, japanese clothes, floral print, headband, hair ribbonSchool uniform:
school uniform, sailor collar, long sleeves, serafuku, red neckerchief, hair tubes, white hairband
Description
FAQ
Comments (8)
Your lora actually manages to get her red eyeliner right, I struggled a lot with that but kinda just gave up as I was training basically every character lol
Her kimono seems to have a hard time getting prompted though, the exact details of her kimono pattern will be near impossible to get right because of the 4 channel vae but prompting black kimono gets rid of the issue of the kimono coming up in a random color gacha. To get as many kimono details down as I could, I tried training "black kimono, floral print, butterfly print, yellow obi, wide sleeves, red rose, pink obiage (the pink sash on top of the obi), and red obijime (the red rope in front of the obi).
The other outfits seem to work nicely though, especially with her adult version which also gave me trouble.
Her dataset (99 images) was much smaller than for Kagarino Kirie (143 images) and I didn't find as many sprites so it made training her harder. I was also a bit aggressive cropping out hands to avoid screwing up the anatomy knowledge of the base checkpoint. Overall I'm not 100% satisfied but training two versions of the same lora back-to-back gets me quite exhausted.
Sorry about not noticing your own lora. I usually search on Civitai before training one so I don't just do yet another variation of what's already available. I think I just searched for her name so it didn't come out in the results. I hope you still had fun with it :)
As for the kimono, I did have "black kimono" in the image metadata so I guess I just screwed up the outfit prompt in the description. Thanks for letting me know.
When a flux finetune or something comes out I'm planning on redoing training with loss masks so that hands and such won't be too big of an issue. Also I don't mind that you trained them at all, at the end of the day my lora is going to be less accurate as a result of having all the characters under one lora so cross contamination will be inevitable, more specialized character loras in that case will always be superior.
@5e5 Oh, being able to add masks to the dataset sounds amazing. Do you mind sharing some info about it? I use LoRA_Easy_Training_Scripts so I don't know if it's possible to do with it.
@happy_tentacle Here's a rentry guide for it
https://rentry.org/d2ckzxmq#loss-masks
I tried looking up derrian's github to see if anyone talked about masked loss training but it doesn't seem to have even been brought up before so I don't think it's possible oddly enough. Given how the advancement is at least 6 months old now I'd have expected it to be added, just last week we found out debiased estimation loss is possible as an argument for training too so that disconnect is odd to me. Try asking him on github if you're interested, I'd also like to use easy scripts for loss masks myself lol
Also it seems that using loss masks dampens learning so you'll need to train for longer. Also be careful with issues involving composition, you wouldn't want your lora to be an unintentional dutch angle lora lol
@5e5 Thanks for sharing some of your knowledge with me. Much appreciated 😊
Alright so I accidentally deleted all my tomls for easy scripts and decided to update, and as I was using the metadata to rebuild it, I found out that there is in fact a Masked Loss setting in easy scripts. It's a checkbox under Optimizer Args, bottom right of the drop down menu, between Max Grad Norm and Huber Param. Then under the subset args section you should see "Masked Image Dir" under "Input Image Dir."
Happy experimenting
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.