These models' art style is best with semi-realistic, though they capable of anime style as well. I use them to create a very wide variety of things, from highly detailed armors to cute anime "girls." These are softcore models, I do not use them to generate hardcore imagery, and in the few tested of it I've done, it did not turn out well.
MK 2 of the model has been released.
Description
This is a reliable model I've been using for quite a while.
I've included a wide variety of sample images, so hopefully they'll give an idea of how diverse this model is.
Uses AnythingV3.0, SD 1.5, Zeipher F222, Trinart2 115000, and Trinart Characters 19.2m. It's a pretty old mix, but it's been real damn gold for me.
It seems people have been having trouble recreating my prompts, and I'm not sure why. It's possible I changed some settings that funk it up without realizing.
FAQ
Comments (16)
Oh man, this is gold. But the file size is quite expensive 😂. I mean, my internet speed is under 1 MBps 🤣
I feel you, it took me like and hour and a half to upload this
But my download is complete now, about 5 hours or so 😂
Hope it ends up being worth the download :)
I hope so. Since right now I can only experiment with Google Colab, I don't have the proper hardware 😅
@aimer if you use Colab you only need to cloud upload this to your drive and use the ckpt from there directly. Use this: Model DL.ipynb - Colaboratory
Thank you for the advice 😁
This looks really nice, but could you also provide a ckpt? I don't know how to convert safetensor models to diffusers, it only works with ckpt files.
Filetype has been switched for ckpt
Did you use GPT to create the promtps?
2 or 3 of the displayed images were from that, the majority of the prompts were made by me though.
yeah, I noticed it by the style, thanks!
Please tell me, what is GPT? thanks
Ive been trying this out and ive been having some trouble replicating the sample images. What VAE did you use? Thanks in advance!
This is copied/pasted from when I tried to help someone recreate my gens on an embedding I released:
Your results could change depending on which version of the webui you're using. What GPU you have can also quite change the output (I found that out when I upgraded from a 2060 to 3090 - I was unable to recreate images, with some being quite different). I don't think the specific VAE has a major impact on the output, but my most commonly used one is the NAI one.
If you're a recent version of the webui, make sure you check 'Use old karras scheduler sigmas (0.1 to 10)' and 'Upscale latent space image when doing hires fix' in the settings. If you've got the latest version... well, highres fix had some big changes to it, and it rather greatly changes how things come out - I don't know that there's a way to fix that just in the settings, so you'd have to revert to an older version.
If you've got the latest, you can do 'git checkout fd4461d44c7256d56889f5b5ed9fb660a859172f' to have a slightly older version with the old highres fix. I've confirmed that the pics will generate properly on that (assuming you've made the settings changes I mentioned) during my tests.
To add to that, some of the pic may have been generated on a much older version of the WebUI, and even I can no longer quite completely replicate them.
Much appreciated! Ill give this a try