Use original text encoders alongside NF4 transformer model (mix and match)

#3
by NielsGx - opened

It seems the text_encoder_2 file has also been quantized to NF4 (1.29GB vs 2.78GB)

Can I mix-match and use the original text-encoder files, and use the NF4 transformers from here ?
Same question with the VAE (159MB vs 169MB)

I believe the text encoders are ran before the model, so using the full ones should net higher quality without increasing the max VRAM usage.

NielsGx changed discussion title from Text_encoder_2 size to Use original text encoders alongside NF4 transformer model (mix and match)

Sign up or log in to comment