_v1 behaves quite differently from the prior iterations
It produces a Very Cool "rough sketch" effect when run at 0.45x and 3 steps, however, which is super useful on its own
will report more Funnies
Hm, it has a fun interaction with this strange Thing:
https://civitai.red/models/2772734/anima-wild-turbo-lora-personal-mix?modelVersionId=3121986
Under normal conditions, this particular 'tinkered' or 'butchered' (down to viewer interpretation) turbo LoRA produces a lot of sawtooth patterns in the image which get reinterpreted into different shapes based on text conditioning.
However, 3 steps is not normal conditions, and so it instead ends up producing surprisingly vivid imagery in those 3 steps.
From 0.05x to 0.6x it serves roughly as a "draft quality" slider; above that, it starts compromising output diversity once again. Starts artifacting above 1.2x under normal circumstances..
Fortunately, I happen to already have tested similar things Outside Of Normal Circumstances, so you can actually crank it up to a 2.0x total and do a single small sampling step without noise
superturbo_v1 / wild
0.45x / 0.0x:
0.45x / 0.42x:
0.45x / 0.42x, plus a single noiseless start-7-end-8 step at (0.45x / 2.0x):
If anyone is curious as to "the value of noiseless refining steps", then the primary benefit is Consistent Output; the output remains mostly the same, only shifting slightly as the VAE encodes the input image slightly differently.
This is very handy in an image editing setting, e.g. brushing a REFINE SCENE PLS output which could realistically already be pre-computed in the background, with a silly 'enhance this scene' brush whose output you have to confirm (e.g. by pressing Enter) such that the next output could be computed..
It's nice to not be swimming in pure 'everything is always changing' AI chaos, and have some semblance of deterministic output.
example of noiseless Stuff:
input image:
output image:
input image:
output image:
Some cute white birds appear from the white pixel attack, but the rest of the image stays comfortably (though not entirely) unshifted
looks really nice. the new method I tried (continuous time distillation, https://arxiv.org/abs/2605.06376) makes model smoother beyond target step, so can work in step 3 I guess.
Hm, it works quite funkily at lower strengths with krita-ai-diffusion, as it usually results in weird intermediate sampling steps when using low 'strengths' and a low base amount of steps (in this case, 3 steps)
https://www.dropbox.com/scl/fo/j5mran1pfoclzy8sgp8bv/AE7lobpaLwPDSYf_U854MEw?rlkey=dwwv5hgb4htrtqduecrpcm32u&st=83upck8q&dl=0
since the interface of this website is kinda questionable for Large images, have a dropbox folder
the initial draft is very fuzzy, but then when you re-run the same 3 steps at lower sigmas, it resolves much cleaner imagery even without having to flip-flop models
would be better with easier sampling steps adjustment inside of krita, but it still kinda works
I see. rather sorryhyun/anima-turbo-4step may work better since it was trained to generate not only with 4 steps but perturbed steps.
The 4-step is better for mixing with other wacky alternative turbos for a very nice 8 step txt2img output, but this particular 2-stepper is rather unique in how it lets you start with a rough draft image which you can very easily Manually Edit, then refine into particular detail.
txt2img has one set of 'feature requirements' for high performance
img2img, and txt2img->image-editor->img2img -- has quite a different set of desireable features