circusmix_v40.safetensors

6.62 GB · SafeTensor · fp16 · full

Abonnement requis

<circus4-xl>

These checkpoints are for loras mainly, hence the style flexiblity, usually I test my loras on these, check if they look somewhat like the training material and I don't like to switch checkpoints around, so rather have an all in one checkpoint. It's still kinda flexible, but has a realism bias. It's now at a spot where it's mostly for realistic characters. The reason I upload them is to avoid ppl asking me how my stuff look like that. Use this VAE: V7.1: Took v7 and added a VAE, and replaced the text encoders (clip_l and clip_g) with the ones extracted from IL XL 2.0 after someone left a negative review complaining about the encoders. I wasn't aware you can just replace those. Keep in mind that those encoders and VAE might not be fully compatible with this version of the UNET, but the small amount of testing I did sofar show no signs of issues. (If you do, let me know.) P.S. I always wondered why this checkpoint was ~300 mb larger than other models, and now it's at normal size. Maybe it did have double entries or whatever, I'm not going to pretend I understand any of this. V7 (no VAE): Exact same checkpoints used as V6, but different settings. This one is very sensitive to styles and prompts, reproduces faces more easily and more accurately. The reason I made this is because I've reached the limits of the pervious one and needed to upgrade, didn't think it would be this easy but!!! it still has the issue I made this in the first place, discoloration of the skin on flat color cartoon characters when trained strong enough, noticed this with some disney characters. Overall flexibility and accuracy improved, managed to mitigate the skin coloring to some degree, but the checkpoints that are good with realism will never be too good with the more flat cartoony look, sad. I don't use any quality tags with it, not in prompts or negatives. V6 (no VAE): It's a mix of and (i still have 's permission to share merges of his models -I asked, just incase-). Flexible with styles and does better realistic characters. -I use Swedishfish because it has a secret ingridient and alex makes the other half of that secret sauce in his checkpoints- V5 had issues with face consistency, especially characters with realistic face proportions, at first I though it was the loras, later I found out it was actually the checkpoint since V6 can keep the faces really consistent without messing up the eyes or mouth (those 2 were the most common problems), and too often it just generated children with some realistic loras which was very annoying! It's still good with animated styles. V5 (no VAE): It's a mix of , and (my man gave me permission to share merges of his models). V4 had issues, like general lighting (too bright or too dark images) and discoloration with adetailer (I know thats a more common problem, but that occured more often with V4), This one has a pleasant base style while flexible and can handle realistic characters. V4 (no VAE): Is the same as V3 except it's a mix of tOOnBabes and Illustrious 2.0 (50%-50%). It's been up for days now, but didn't publish it. Someone asked me very nicely to release it so made some showcase images. V3: People were asking about this one for the style (especially with the Overwatch characters), so I'll share it. Wanted something flexible thats closer to original styles with loras without the forced anime to see if trained loras actually came out the way I wanted. It still does anime well, but atleast it won't force it on loras or when using style tags. Still not style neutral but good enough I guess. It's a mix between my tOOnBabes and Illustrious 1.0 (50%-50%). - I still don't know what I'm doing, just mixing stuff and see what comes out - V2 (Forgot to include VAE): Was playing around with some merges, so V2 has a different base style, is still flexible and anime is still there but not as strong. ps: I don't hate anime, there's just too much of it, not all of us are fans. psps: Maybe later i'll post some more showcase pics, don't think they're that important anyway. V1: Wanted to create a merge of illustrious that's not too anime heavy, I think I succseeded.

SHA256

1BE6E7324F5BF42C55A14DCB9DBC92F64ECEE8B52B91E7EF31587FC14D562ACE

ID de modèle ModelTrends.ai

#19719