Type | |
Stats | 26,703 |
Reviews | (8,051) |
Published | Mar 3, 2023 |
Base Model | |
Hash | AutoV2 8AC3E79E96 |
TLDR of everything below: "Use absurd2 version if you use HiresFix. Use normal version if you don't use HiresFix. When absurd2 is unstable, use withContrastFix version. When it's too black, use normal version. Use included VAE (nai). Consider using CFG Fix extension, most example images used it."
If you happened to try out this mix, please leave a review, because I'm curious about how it performs with other people prompting styles!
There's currently no version that is better or worse than the rest, different versions are just different style-wise, or have a different "detailed but possibly messy : stable but possibly boring" kind of balance, to which every mix approaches differently. Check out example images of different versions of the mix! (those are not cherry-picked and are directly comparable between versions)
Also the "How" section explains some differences between the versions.
Why.
Out of curiosity I've made a model out of the models I've been using, just for personal use, but the result is so good that I decided it was worth to share the nice stuff with more people!
I really liked the rich crispy details of YOZORA, the crazy semi-realistic capabilities of Xtracolor, the artstyle of AOM, the creativity in details of Counterfeit and the faces from MeinaMix and Korean 2.5D. Also added the uiouiouio's models because I liked the results overall and those were close to what I was trying to achieve.
So, Nyan Mix ended up being able to produce very colorful images with high contrast and rich effects in semi-realistic styles, but at the cost of making environments too abstract (unless you're using CFG Fix!), and being very dark on average by default (when using "withContrastFix" versions).
What.
The recipes are described in the versions' changelogs.
From my initial testing, it's best to use "normal" or "withContrastFix" version as a general-purpose model and "absurd2" for dynamic scenes with darker backgrounds.
The "absurd" version is much more detailed compared to "normal", while "intense" version is somewhere in between. Doesn't seem to be as unstable as I expected when used with certain parameters (namely Highres Fix), but amplifying it even further breaks it, as well as using it with some other parameters. Might become unstable enough when used with LORAs, but I haven't tested that.
The "absurd2" version is very dark compared to "normal", but the contrast is great. The "withContrastFix" versions are somewhere in between.
You can switch around the version tabs a little bit lower here on the page - the example images are made so that you could directly compare the different models' features.
Note: 230301 version is more 2.5D-ish, while 230303 versions are more 2D-ish by design.
Comparison between versions (will be updated with more examples soon): https://imgur.com/a/JvSQqUV
How to use.
Maybe it's not the best configuration (I'll be happy if you'll discover a better way), but that's what I used to generate most of the example images:
Sampler: DPM++ SDE Karras
Highres fix: Enabled
Upscaler: Latent (nearest)
Upscale by: 2
Sampling steps: 20-50
Highres steps: 15-25
(generally 20+15 steps is already good enough)
Denoising strength: 0.6
CFG Scale: 7
Clip skip: 2
Resolution: basically everything works well - both portrait and landscape aspect ratio capabilities are good, both low-res bases (200x400) and high-res bases (896x1344) are also good.
VAE: Included with the "absurd2" version, it's the same file that Counterfeit/OrangeMix/RefSlave/etc use. Other VAEs might degrade eyes and introduce severe unintended "chromatic aberration" effect.
Consider this!!!:
https://github.com/mcmonkeyprojects/sd-dynamic-thresholding.git
CFG Scale: 15
CFG Fix: Enabled
Mimic CFG Scale: 7
Top percentile: 95%
Mimic Scale Scheduler: Half Cosine Up
CFG Scale Scheduler: Half Cosine Up
Minimum value of CFG Scheduler: 3.5
Power Scheduler Value: 4
(Most of the example images were made using this)
Alternatively:
Sampler: DPM++ 2S a Karras
Highres fix: Disabled
Resolution: preferably no less than 800*1200
CFG Scale: 7
Sampling steps: 10-30
Alternatively:
Sampler: DPM++ SDE Karras
Highres fix: Disabled
Resolution: preferably no less than 600*900
CFG Scale: 8.5 or lower depending on resolution
Sampling steps: 15-35
Plans.
Will try to reach for better stability and even better backgrounds.
The next version will be more 2.5D-ish.
Will try injecting different styles to see if something will look like it belongs here.