Update VAE
#4
by dn6 HF Staff - opened
@YiYiXu correctly pointed out that the VAE here is the similar to the SDXL VAE, but uses 16 latent channels and does not use quant conv and post quant conv operations. We can reuse the existing AutoencoderKL object for this model.
dn6 changed pull request status to open
dn6 changed pull request status to merged