lob10sterstyle, androidseams
Source: https://civitai.com/models/698073/loblob10ster-car10-alt-artstyle-bonus-android-seams
My best quality LoRa yet, and I trained it myself.Keywords: lob10sterstyle, androidseams (that's right, BONUS TAG; does what it says on the tin)Recommended Weights: 0.8-1.1 (Finally something reasonable. No underfitting this time, take that PixAI!)Happy generating.(We can finally give these androids a good ending for once!)If you like what I make, please do leave an upvote or a tip! It'll help a lot towards being able to make more LoRas in the future.Known Issues (V1):Green Triangle on Chest with "androidseams:" Very obviously from some of the images in the dataset (you can even find which manga it's from), will be cleaned up in the next version."androidseams" appearing over clothes: In my brief stint of testing, I found one case of the seams being applied over clothes. There seems to be a conflict with seams and clothes since the seams are most prominent when there is more skin shown. I hope adding more images to the dataset amends this.Night Club Background: Again, another artifact from another specific manga. Will be tagged and fixed in the next version. Not a big deal most likely, but putting "strip club, nightclub" in the negative prompt works in the meantime.Line Clarity: Maybe I'm just nitpicky or imagining things. We'll see if more data will improve this.Negative Space: Sometimes the spacing and position of the characters can feel... somewhat off? Like a full body view will place a character slightly to the right for some odd reason. Will try captioning images with "cropped" because I heavily cropped many images. Hopefully that fixes things? (If you're reading this and have an idea, please do share!)Development Log:Trained my first LoRa from scratch WITHOUT Pixai's (unfortunately limited) LoRa trainer.This was actually a test LoRa to see what processing manga panels would be like for dataset building. I also wanted to test how just a few well-processed (22) images would work for a style LoRa. Turns out, the results are surprisingly good! Quality over quantity holds true.I also was a bit ambitious and tried to incorporate two activation tags without definitively knowing how. I was pleasantly surprised by how well the LoRa picked up the android seams, considering they're just lines.There's so much more data I can add. I might consider slipping in content from the main too in a future iteration. I'm excited to perfect this... possibly.
