Update 1/12/2024
_
v4 is an upgrade on v3, featuring on reduced meaningless random objects and color, better compositions, better eyes while maintaining most of the good features of v3 like nice color and light effects, obedience...
V4 has baked VAE (which is called V1.5prunedVAE or animevae in some other places) so you don't have to load extra VAEs
_
For better face details please enable ADetailer
Recommended parameters:
· Base image resolution: 960*640 (3:2), or 960*540 (16:9) vice versa
· CFG Scale: 7
· Clip Skip: 2
· Highres Fix: use latent upscale (latent upscale method is bilinear, on comfyui you need to select it in Latent Upscale Node, but on webui it's just the default highres fix method just called "latent") to upscale the base image by 1.6 times
· Upscale: use Ultimate SD Upscale with Controlnet Tile (My Prompt is More Important mode) enabled to upscale the highres fixed image to 2K, 4K or whatever
_
Update 4/29/2023
betterMix3.0_Anime and betterMix3.0(with_cetus)_Anime is added.
3.0 is a general improvement on 2.0, focusing on better eyes, anatomy (and a bit better hands) while maintaining most of the features of 2.0 like wonderful color and light and shadow.
3.0 (with cetus) mixed cetusMix inside. Compared to 3.0, it sacrifices some color and light effect in exchange for better anatomy.
This model is SUPER COMPATIBLE with other "small" models like LoRAs (LyCoRis, LoHA, LoCon) and embeddings. I highly recommend using this model as a base for other models' mixture.
All the example images provided here are using hiten1 and EasyNegative these 2 embeddings.
The link to hiten1 embedding:
https://civarchive.com/models/15430/hiten-style
The link to EasyNegative embedding:
https://civarchive.com/models/7808/easynegative
-
This model is made by merging PastelMix, Anything3.0 and OrangeMix together.
I mainly use this model to cooperate with the "hiten1" embedding (textual inversion) I made, the link to it is above.
Suggested basic parameters:
Sampler: Euler a & DPM++ 2M Karras
Sampling steps: 20
CFG scale: 7
Recommend to use this model with highres fix on.
Suggested highres fix parameters:
Upscaler: Latent
highres steps: 20
Denoising: 0.6
Upscale from 640 360 to 1280 720
I've also uploaded this model to hugging face, link is below
https://huggingface.co/SakanakoChan/BetterMix2.0_Anime
如果可以关注我的bilibili频道的话我会很感谢的！
I'd be grateful if you can follow my bilibili channel!
Description
actually 2.0 is the first version, I call it 2.0 because it sounds better:((
FAQ
Comments (18)
作者，有没有huggingface链接啊，civitai下载太慢了
在description最下面我加了hugging face的链接
@SakanakoChan 我看到了，感谢
用idm应该可以满速
can we download this in mac 2 and put this in diffusion b
I don't have mac so I'm not sure if it'll work on mac...
wait i just checked diffusion bee's documents and i guess you can just download it and import it to diffusion bee and it'll work
Is it possible for you to make a pruned model?
Sorry I am not an expert on making and pruning models so probably no...
nvm I've added a fp16 safetenser version
@SakanakoChan Thanks!
Idk if I did something wrong but every time I generate something the image is desaturated. I even tried using the prompts on the examples and it still the same, and yes I downloaded easynegative and hiten1
did you use highres fix? That’ll fix 90% of problems. You can try using the same parameters as what I wrote in the description
And maybe the reason for generating desaturated images is… not enabling VAE?
@SakanakoChan how do you enable VAE?
@SakanakoChan nvm I fixed it I didn't install hiten properly still don't know what VAE is though. Thanks for the help!
@onionsoup Glad to see you've solved the problem! VAE is a stuff to fine-tune the checkpoint model's style (like color and lines), by default you can switch your VAE in webui settings > stable diffusion, or if you wanna move the VAE drop down list to the top left of your webui, you can go to settings > user interface, you can see an attribute called Quicksettings list, and add sd_vae there (just as what the descriptions above Quicksettings list say).
I mainly use the basic vae for stable diffusion 1.5 which is called final-prune for betterMix_Anime, you can try other VAEs as well.
very beautiful!