Conversation

Notices

1. Embed this notice
    (mint@ryona.agency)'s status on Thursday, 05-Oct-2023 07:21:09 JST 
   Tired of Taba, it's Frau neighbourhood now.
   • Embed this notice
      (mint@ryona.agency)'s status on Thursday, 05-Oct-2023 07:25:51 JST 

     in reply to
     Had to bake a LoRA myself. Only ~30 images suitable for training on boorus, so had to compensate with screencaps from a 720p TVRip I had laying around. Probably shit for everything but basic poses, 8th epoch seems to be working best.
     frau-000008.safetensors
   • Embed this notice
      (mint@ryona.agency)'s status on Friday, 13-Oct-2023 23:32:48 JST 

     in reply to
     Doubled the dataset with images from Pixiv, it's at 205 images now, and cleaned up duplicate eye/hair tags. Retraining it with the same default settings, but with net dim 64 and resolution of 576. Sample of the very first epoch looks a bit more promising than last time. Should have tagged her school uniform as well, but alas; it appears in only a few chapters/episodes anyway.
     last_20231013142454_e000001_01.…
   • Embed this notice
      (mint@ryona.agency)'s status on Saturday, 14-Oct-2023 01:46:57 JST 

     in reply to
     Not noticing that much difference with larger epochs yet.
   • Embed this notice
      (mint@ryona.agency)'s status on Saturday, 14-Oct-2023 05:10:56 JST 

     in reply to
     Hands still fucked, either dataset issues or the mix I'm using simply isn't playing well with them. Could be touched up separately, so not that much of an issue. Don't know about overfit, epoch 7 still seems to work fine on 1.0 strength, but plays better with other loras at ~0.8.
     00043-3108233785-_lora_frau-000…
     00042-3108233787-_lora_frau-000…
   • Embed this notice
     Ethical Pedophile (blinkrape@posting.lolicon.rocks)'s status on Saturday, 14-Oct-2023 05:17:11 JST Ethical Pedophile

     in reply to
     @mint hands is one of the hardest shit for AI to do right. idk what you're using but if its stable diffusion, most people use a dedicated LORA trained on hands to help with that shit. Also just train it on lots of good hands, with specific trigger words.
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Saturday, 14-Oct-2023 05:19:56 JST 

     in reply to

     • Ethical Pedophile
     @BlinkRape There wasn't much training material with her to begin with, I was really scrapping the barrel. Haven't looked up hand loras yet, there's a "bad hands" negative embedding I have laying around, but I think it might just be making matters worse. Highres fix fucks it up as well even if the first lowres generation had them looking fine.
   • Embed this notice
     [TDK] Digital Cheese (digitalcheese@ryona.agency)'s status on Saturday, 14-Oct-2023 05:21:13 JST [TDK] Digital Cheese

     in reply to
     @mint >Hands still fucked
     >Better hands than what I can draw
     
     AI W
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Saturday, 14-Oct-2023 05:23:17 JST 

     in reply to

     • [TDK] Digital Cheese
     @DigitalCheese I had to fix the left hand on the first one since piping it through highres fix turned it from a normal-looking curled one into some broken backwards-curling abomination. Just upscaled the first result and copied the hand onto final.
   • Embed this notice
     Ethical Pedophile (blinkrape@posting.lolicon.rocks)'s status on Saturday, 14-Oct-2023 05:24:19 JST Ethical Pedophile

     in reply to
     @mint hi-res is weird. It's basically just running it through an upscaler, but that still gives you options to tweak how much it can change the image if I am not mistaken. That setting has to be set relatively very low.
     
     like 0.3 or lower.
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Saturday, 14-Oct-2023 05:25:30 JST 

     in reply to

     • Ethical Pedophile
     @BlinkRape They changed how it works a while ago, used to be more faithful to original image in older versions. I'm using ~0.35, seems to be the sweet spot.
   • Embed this notice
     Ethical Pedophile (blinkrape@posting.lolicon.rocks)'s status on Saturday, 14-Oct-2023 05:26:47 JST Ethical Pedophile

     in reply to

     • Ethical Pedophile
     @mint You could try generating an image and then running it through img2img yourself with the denoising turned down, or see if tweaking the denoise strength lower, under the hires fix category, helps.
      likes this.
   • Embed this notice
     Ethical Pedophile (blinkrape@posting.lolicon.rocks)'s status on Saturday, 14-Oct-2023 05:26:48 JST Ethical Pedophile

     in reply to

     • Ethical Pedophile
     @mint
   • Embed this notice
     Ethical Pedophile (blinkrape@posting.lolicon.rocks)'s status on Saturday, 14-Oct-2023 05:29:03 JST Ethical Pedophile

     in reply to
     @mint if you haven't, check out this guys channel. He has a lot of informative videos about various topics related to SD, and Auto1111
     
     His controlnet videos are pretty good. I stopped messing with this shit because it was too frustrating on my GTX 1070 which is too slow for my taste
     
     https://www.youtube.com/@sebastiankamph
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Wednesday, 18-Oct-2023 21:27:53 JST 

     in reply to
     Retrained once again, this time back to 512 resolution, but with adaptive Prodigy optimizer. Looks like it plays better with special needs models (AOM3 absolutely fried the previous DAdam versions, this one doesn't look any worse than basic prompts without any loras) and when mixed with other loras.
     https://civitai.com/models/167366/frau-kojiro
     00147-3149967507-_lora_frauprod…
   • Embed this notice
     Adachi (adachi@pleroma.adachi.party)'s status on Wednesday, 18-Oct-2023 21:51:50 JST Adachi

     in reply to
     @mint
      likes this.
   • Embed this notice
      (mint@ryona.agency)'s status on Wednesday, 18-Oct-2023 21:52:59 JST 

     in reply to

     • Adachi
     @adachi I look and act like this.
   • Embed this notice
     Kirino Kousaka (kirino@seal.cafe)'s status on Wednesday, 18-Oct-2023 23:05:42 JST Kirino Kousaka

     in reply to

     shwag, very nifty

      likes this.