NightVisionXL_0.8.1.1 Public Release Notes

I've been working hard on a new training for NightVision! This new update has received multiple targeted tunings aimed at producing more stylized and aesthetically pleasing output. It has also received the first round of anatomy tunings which is targeted at producing more realistic humans, male and female both, including better and more varied faces and races, better overall body coherence, better facial expressions, improved hands, elbows, knees and toes. Genitalia is more realistic, though not perfect. The model is now more "penis aware" in that if you prompt for a naked person (man or woman) with a penis, it's gonna try to do it, to varying degrees of success. It also can do more varied poses for naked women as well, and is better aware of both male and female sexual organs. It is definitely a WIP and it won't take you long to get a ken doll dude, but it's very much a balancing act trying to get the model to realize what a penis looks like and where it goes without putting it everywhere. The model is also much less strict on it's style of output, and is now much more of a generalist model. It still outputs beautiful photographic style output, but it can also lean into other styles such as anime, painting, sketching, etc.

Aesthetically this is a big step up from previous versions and can produce really beautiful scenes with very simple prompts, however I've also been training the CLIP model as well, so you can expect prompt coherence as good as or better than any of the other top SDXL models around today. NightVision speaks in natural language, prompt it with words and sentences and it will surprise you with how well it understand your prompt. It's pretty good with text too!

ChangeLog 1/19/24

* roughly 40 hours in train time over 7 targeted trainings:

* around 1500 total hand selected AI generated images chosen for their quality and aesthetics (sourced from Civit, MJ, Dalle3, Playground and Stable Diffusion).

* 600 photos randomly selected from a license free amateur photography dataset from HF

* around 800 randomly selected images from LAION-POP dataset, though I did remove several dozen images that I didn't want making it into training as they would have been problematic (like weird yoga poses as an example).

* targeted training tokens include "skateboards, surfboards, clarinets, saxophones, guitars, pianos, smoking, bicycles, vehicles, airplanes" and many many more!

* added training human-anatomy-nv7131-001

* Backmerge! 7.14.0 was a bit too nudez happy, so did a back-merge with 7.13.1 (pre nudes) and 7.9.1 (to help bring back some of the color neutrality and re-focus the photography style as the default output)

Known Issues

* noting some issues with patina on some photography tokens like analog and polaroid. I've managed to curb most of it, but the behavior still pops up from time to time

* Penises are still a WIP, especially from behind.

* Hands are vastly improved, but still can be frustrating sometimes. Inpainting hands works really well though, and can add incredible detail if done properly.

* crowds and faces at distance - coherence for crowds is getting pretty good and faces and eyes at a distance are much better than they used to be, but still noticeably lacking detail.