Work Notes Week of 231211 (Patreon)

Matoi Ryoko

Two Ryoko LORAS - I used both SetsuArts and Yesmola's LORAs for this set. SetsuArt's LORAs work well with 'western' concepts and LORAs, much less so for bondage, while Yesmola's LORA works out of the box with the new RichyRichMix model. I've said this before, but there is a big difference in the western and anime artstyles that result in a clash.

The Dangers of Multiple Outfits - It is difficult to train a LORA with multiple outfits due to the bleeding that occurs. The two outfits will start merging if tagged improperly. Ryoko's Senketsu outfit needs help with 'underboob, suspenders, microskirt'. Those are the things that make it stand out from the normal outfit.

Perfect_Anatomy - Tends to give the 'Mikasa effect', where the subject has toned bodies like Mikasa. Also gives it a shiny, glossy body not unlike the DOA girls.

Upscale Size Affects Quality - Sometimes I am tempted to just upscale at 1.5x and be done with it; the problem is that Upscale Size Affects Quality. Many times an image comes out blurry and unusable because I forgot to set the upscale size back to 2x after the CPU conked out. This happened with one of the Ryoko images and as I was reviewing the final six, I realized that one of the images was below par. Is it a bother when the long wait times results in unusable images? Sure, but the quality differences in 2x upscale and 1.5x upscale are too great to ignore.

Use Interrupts If Image Isn't Genning Well - A big downside of upscaling at high denoise & high dimensions is that the result can get a little funky. Extra hands, unintended effects etc... One of the Ryoko pictures had this "door opening to cell door" effect going on and I really wanted to keep that, but high dimensions made it go away. You should resolutely interrupt the generation if things aren't going well on the preview. Otherwise, that's a lot of time lost generating a picture that needs a lot of Work or simply doesn't work.

Kyouka Uzen

New Model, New Work - It was difficult to work with RichyRichMix at first, a lot of the images came up scuffed (a recurring issue) in that the faces are 'burnt' or pixelized and make for really bad previews. Upscaling fixes things, but I would rather have 'preview' images that are good, that gives me the confidence to spend time tweaking and upscaling it. 'image burn' is a recurring issue with newer models it seems.

EASYNEGATIVE WAS THE CULPRIT - After a lot of trial and error, I finally gave up and just copied the prompts off the Kyouka LORA page. I got a perfect result, and I started to look into the differences between the author's prompting method and mine. Turns out that the only real difference is in the negatives tag, where I use a really old embedding called 'EasyNegative' which is a whole bunch of negative tags together. I removed that and image quality drastically rose. It really hit me then that 'You need to be able to go back to zero to go forward.' If I didn't try to completely change my style, I would be stuck thinking that the Mato series of LORAs were unworkable, despite the preview images showing the contrary.

Upscaling Dimensions Matter - I never noticed this before, but the upscale sizing matters. At constant denoise/seed it does change the image a tiny bit, sometimes adding fingers, sometimes adding lip gloss, I am unsure what causes this, but even when upscaling, dimensions matter.

Fingers on Tentacles - A strange error! Why does the upscaler think that tentacles have fingers? My hypothesis is that tentacles are grabby, which makes the model think of hands, so tentacles have fingers. Tada! Flawless logic! NOT!