Update: CLIP fix - now we get even more accurate and more beautiful images (Thx to the Explanation about the Model Toolkit Extension by @ndimensional)
I merged it a time ago and don't know anymore which models went into this one (Art&Eros and HassanBlend i can remember), but its my go to Model for almost everything. It was aimed at photorealism but it can do a lot more. Thought maybe it's useful for the public.
U need to use a VAE!!!
I use standard VAE: vae-ft-mse-840000-ema-pruned
(added VAE-file for download to the CLIP Fix Version)
For Automatic1111, put the VAE into ...stable-diffusion-webui/models/VAE folder and make sure you set it in the settings.
Description
FAQ
Comments (6)
Think you need to fix the broken CLIP. Quick test: use woman in the positive. If you dont see a woman then there is something wrong
Dont know what u mean, i get the result even with a single keyword. It depends on the seed and also on the weighting of the keyword. So no Problem here i guess
@epinikion It's working (and also great results). But with tensor CLIP problems you can get dropped chunks distortons, hands, faces, embeddings using, etc.
@Cyberdelia Okay, never heard before. Wouldn’t even know how to fix that. Never had problems with embeddings. If you know how to fix or improve that, let me know.