TWO New Fizzogs for You - Updated! Detailing, Weirdness Enabling Flux Facelift Experiment

8/3: I added a second version, little cheerier output, maybe a hair easier to use, less grain, slightly more of a rendered look, and not quite as gore prone. Probably still will seem like darn near the same LoRA to others.

Fizzog=face. Consider this a facelift, detail enhancer, and enabler of weird/childish scenarios. If you are familiar with the past LoRAs I have trained, this will do pretty much everything in one, minus the characters, or will work as an excellent detail enhancer (if you have my tastes, and I certainly do). Mild and pretty to wild, insane cursed, your pick.

Works for anything, particularly designed to allow more flexibility and give things a facelift, but some random subjects it likes- vintage photos, darker scenes, fashion, cars, multi-subject prompts, fighting, easy punching and kicking in the face, fleshy blobs, comedic unrealistic wieners, comedic gore, CRT screenshots (finally, scanlines and bezel at the corners by default!), better understanding of putting objects in mouths, digestive issues, statues, floral scenes, landscapes, etc.

Use

Gritty Version

Optional trigger:

a n3wf1zz0g4y4

Start at around .7.

Can use it higher but may get some artifacts or undesired warping (pending on prompt). Higher strengths may give you ridiculous NSFW if you ask for it (fake comedic phallos and the like). Sometimes cranking it to 1.1 or 1.2 is good. Makes it all glow in some cases!

Lighter Version

Optional trigger:

a n3wf1zz0g4summ3r 

start around .8-.9

Both versions work great with Krea too!

What is this?

Lipstick on a pig of noise. Trained on a large data set, mostly comprised of old data set images, my favorite outputs, and the kicker- many “problem” images I had to cull from past LoRAs, somewhat based on the noise/shapes training on them led to. Can I leverage the problem items and a huge variety in a data set to make Flux more flexible/better to my liking? I wanted to fight with overtraining and noise to see what happens.

I ran the initial set for 100 repeats each over 10 epochs, tested, updated the data sets to address what I didn’t like, repeat at 5-10 epochs with lower repeat counts, then repeat again, and again. Swapping images in and out each time to sculpt the noise. Far beyond what your normally would. In the end, this was 60 epochs over 34,400 steps!

Think of it as a sandbar at the beach…. You walk out further and are in over your head, but swim out further, and you can stand again. Same thing applies here- overtrain to the point that things start getting messy, swap/update data sets, and the noise on top (your grains of sand) trains.

This was absolutely painstaking, and IDK if it was worth it. I attempted this several dozen times with different tag approaches. What worked far, far better- Janus pro tags, and don’t even look at them, just stick a consistent trigger in front. Florence tagging made things a huge mess. If I were to do something this size ever again, I would use Janus again, but remove all punctuation and capitalization.