Following on from RCNZ Dumb Monkey, this time I smashed together a pile of photorealistic models and the awesome noise offset model to make another mix!
The SD-v2 (840k) VAE is baked into the originals, and all are pruned to fp16
Description
FAQ
Comments (12)
Great mix and great quality of the renders.
Can you also add the no-baked-vae version ? Which vae was baked inside the model ?
This has the sd-2 VAE baked in (vae-ft-mse-840000-ema-pruned) - I'm not actually sure how to unbake it tbh, but happy to oblige if I find out how!
@residentchiefnz Thanks. Not sure if this well known script can be of any help:
https://github.com/arenatemp/stable-diffusion-webui-model-toolkit
It says:
"Any merge between different VAE's will result in something broken. This is why the VAE in the AnythingV3 checkpoint produces horrible outputs, and why people have resorted to distributing and loading the VAE seperatly."
Another useful script:
https://github.com/Akegarasu/sd-webui-model-converter
convert to precisions: fp32, fp16, bf16
pruning model: no-ema, ema-only
checkpoint ext convert: ckpt, safetensors
convert/copy/delete any parts of model: unet, text encoder(clip), vae
@ritcher1 I used that toolkit to bake in the VAE, Ill have a play with it and see if it can unbake it too
From experience I know baking in the NAI/OrangeMix vae goes badly, but never had issues with the kl-f8-anime (WD) or 840k (SD) ones
@residentchiefnz Ok, no problem then. I always prefer the modular use of the vaes, more flexible. Thanks anyway.
@ritcher1 no worries!
@ritcher1 - safetensor no VAE version uploaded
@residentchiefnz Thank you ! The v2 is baked-vae or not ?
@ritcher1 yet to create v2-noVAE, gimme a day now I know how to do it :)
@residentchiefnz Thanks !
@ritcher1 That new version is now there for ya chief :)
@residentchiefnz Perfect ! Thank you very much ! Already downloading.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.