Hi there, I played around a lot with stable diffusion marges those days, and from this i decided to make one for me and my tastes.
This model uses Anylora as a base, then it was mixed with multiple anime base models (sorry i can't give you more details because i don't know them my self XD) till i ended up with a model that matched my tastes and that was "different" from other models i've seen.
What is this model objective?
Well the goal was to achieve a good model that excels in portraits, but that can easily be used with good results in any scenario.
It works best with images of girls, but it can produce boys, landscapes, and even NSFW images.
Suggestions:
I've tested quite a while (till my gpu begged for a break) and i have to say that it can give good results in both clip skip 1 and clipskip 2.
The samplers i tested were Euler A, DPM++ 2S a, DPM++ SDE Karras and those have produced good results, but feel free to test. (If you want a more precise setting you can check the bottom part of this page).
The LoRa compatibility is quite good, i tested them a lot, and it perform goodly enoght even with my "failed ones".
The In-Paint is quite good, especially for female faces, but you rarely need to do it if you use hig.res.fix.
[Attention]: HighRes.Fix is not mandatory, but it is highly recomended!!!
As upscaler you can use eighter: "r-esrgan 4x+ anime 6b" or "ultrasharp anime".
Suggested Embeddings
EasyNegativeV2
BadHand-5
aid210
an10
negativefeet
deepnegative
verybadnegativeimages
Examples of settings:
Sampler: Euler A
Steps: 40 - 60
Cfg: 7
Image resolution: 512x512 ; 512x768; 768x512 (it works even with higher ones but there's really no point in sterssing the hardware when you can upscale later)
vae: vae-ft-mse-840000-ema-pruned_fp16
------------------------------------------------------------------------------
Upscaler: r-esrgan 4x+ anime 6b
steps: 10-20
denoise: 0.1 - 0.4
Sampler: DPM++ SDE Karras
Steps: 30 - 40 (it might work with higher ones, but i didn't test those)
Cfg: 7
Image resolution: 512x512 ; 512x768; 768x512 (it works even with higher ones but there's really no point in sterssing the hardware when you can upscale later)
vae: vae-ft-mse-840000-ema-pruned_fp16
------------------------------------------------------------------------------
Upscaler: r-esrgan 4x+ anime 6b
steps: 10-20
denoise: 0.1 - 0.4Description
This version can produce good compositions and has overall better output than the v1 version, but it might need some In-Paint for very complex images.
It is a bit less "2.5d" (even toh it can still produce those outputs), and has improves lora compatibility compared to the V1.
As for version1 it works better for portraits, but can achieve pretty different results (i went a bit overboard with the example images of the post, but those aren't really cherry picked, they are just upscaled, not even inpainted once), it still works pretty good for NSFW content, but now it's less likely (it might still occour) to have a NSFW output if not desired.
FAQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.