Sign In

KNXO's experimental crappy model mix (Anything 4.5 based)

48
495
6
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
155
8
2
Reviews
Published
Sep 27, 2024
Base Model
SD 1.5
Usage Tips
Clip Skip: 2
Hash
AutoV2
9A076E3AA8
default creator card background decoration
knxo's Avatar
knxo

Models:

  • v1.1 Bakedin VAE: KNXO's experimental crappy model mix with fixed Keys and clearvae 2.3 baked in

  • V1.0: KNXO's experimental crappy model mix with original VAE from Anything 4.5 by andite.

  • Anything4.5 ClearVAE FP16: Original 4.5 by andite with baked in clearvae 2.3, key fix and saved as fp16.

  • Anything 4.5 OG VAE FP16: Original 4.5 by andite with original VAE, key fix and saved as fp16.

The last 2 I just added because they can be useful, for generation online simply use the OG file by andite and remember to select clear VAE. Mostly i have added these ones because people kept using my prompts without adding the VAE and end up getting blurry images.

Erm, was playing with Kohya's model merge then with supermerger. A couple days back I made a negative embedding for anything 4.5 So the idea went like this: what if i do a crap lora with my negative embedding dataset and then invert it and merge it onto anything 4.5.

So i trained a literally crappy lora and stuffed it into 4.5 with a couple of my LORAs. Honestly the result isn't that great but it doesn't need as many negatives as OG 4.5.

The result is: horny! In average a slightly bigger cup size, thinner waist and bigger hips. Bias towards making females is even bigger. Bias towards nudity increased by a lot. Dudes look odd! (Might like it if you like slightly feminine bishonen types?)

So I spent about a week playing with it to tweak the block weight and it was likely going to simply rot in my hdd forever so instead here it is.

IT should work ok-ish without negatives or positives, you can still use them and it looked better with my negative embedding