448px (Turbo) Helper

I really like the Krita ai-diffusion plugin, and I like drawing at low resolutions - so I needed a little LoRA to help make the model behave itself at lower resolutions.

Sure, there's the "rdbt 512px" thingymabob, but that one suffers severely from "prints the same picture every time"-itis
So, I guess it was time to make one myself, duh.

now if only I could get comfyui-ppm introduced into said krita-ai-diffusion plugin for the NegPiP cfg 1 negative prompting (which works Better than cfg-whatever 'actual negative prompts'!!) ... that would be nice


useful links:
https://github.com/ThetaCursed/Anima-TrainFlow
handy-dandy super easy portable trainer, that actually works, holy heck(!!)

https://github.com/ethanfel/ComfyUI-LoRA-Optimizer
super useful LoRA mixing/mashing/downscaling utility
surprisingly robust



model description:
448px-Helper:
seems to serve as a sort of cfg-distillation and pixel-ungonkulator at low resolutions
in fact you can achieve increased pixel-ungonkulation and reduced cfg-friedness by loading in the "rdbt_v0.6_cfg_distill_only" LoRA at negative strengths alongside it, hehe

It did its job of assisting in CFG-less generation of images alongside a touch of anima-turbo and that one cosmos2.5-distillation-lora

hey, speaking of which, what if we did The Opposite of what the eggheads say ("don't try mixing DMD/distillation LoRAs") .. what's the worst that could happen?


448px-LazyTurbo:
The star of the show!
Some weird mix of 0.93x 448px-Helper, 0.24x anima-turbo-0.1, 0.10x cosmos2.5...-distilled

What does it do? It generates pictures! They look pretty nice, they 'cap out' around 12 steps at its default 1.0x strength, though additional steps do improve 'new concept LoRA compatibility' as per normal DMD2 business.

CFG 1, sgm_uniform scheduler works well but others also function,
sampler mostly whatever - down to what "visual style" the image has, but euler/sa_solver/er_sde are generally solid choices

Furthermore, it also lets you generate reasonably clean images at lower resolutions!
Which means you can run batches of 4, 6, 9, without it being slow..
This goes really well alongside its Key Feature of "It Generates Surprisingly Varied Outputs"


Due to its bizarre makeup of pixel gonkulation and hard DMD 'stabilization', it does actually function at strengths above 1.0x - some artist styles even benefit greatly from values above 1.0x.
You can crank it up to 3.0x if you so desired, however, due to the inherent nature of 'Distribution Matching Distillation' (DMD) - you will be reducing the pool of valid outputs the higher you go.
In fact, at the aforementioned 3.0x, it wraps back around to 'mostly printing the same image' :- )

cranking up the strength beyond 1.0x also lets you reduce the step counter to something less than 12, like 8, or whatever approaches the theoretical limit of the 4-step cosmos2 distill mixed with the 8-step anima turbo



Can the mixing be improved? Can the lowres-ungonkulation be improved? sure, always
does it work shockingly good here and now?

Y E S

gl hf soldier, go wild!


how it was made:
https://www.dropbox.com/scl/fi/xzojn3jn31ql032huo1i3/448px-data.7z?rlkey=kk5dql8v12yv2spnx7zmw77kt&st=erdyt0w9&dl=1

step 1: generate 512 promptless 448x448 images that get upscaled to a stable (640x640) resolution and refined, then scaled back down to 448x448

step 1b: introduce mild negative prompt bias to ruin everyone's fun (very important)

step 2: let it rip on the unlabelled promptless dataset

step 3: ???

step 4: mix most promising checkpoints until delicious result is cooked

step 5: mix this delicious mixed result with anima-turbo and cosmos2.5-distill weirdly until delicious result is cooked

steps 6 7: (you are here)