Last active
December 26, 2023 07:15
-
-
Save ProGamerGov/70061a08e3a2da6e9ed83e145ea24a70 to your computer and use it in GitHub Desktop.
Replace the VAE in a Stable Diffusion model with a new VAE. Tested on v1.4 & v1.5 SD models
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# Script by https://github.com/ProGamerGov | |
import copy | |
import torch | |
# Path to model and VAE files that you want to merge | |
vae_file_path = "vae-ft-mse-840000-ema-pruned.ckpt" | |
model_file_path = "v1-5-pruned-emaonly.ckpt" | |
# Name to use for new model file | |
new_model_name = "v1-5-pruned-emaonly_ema_vae.ckpt" | |
# Load files | |
vae_model = torch.load(vae_file_path, map_location="cpu") | |
full_model = torch.load(model_file_path, map_location="cpu") | |
# Replace VAE in model file with new VAE | |
vae_dict = {k: v for k, v in vae_model["state_dict"].items() if k[0:4] not in ["loss", "mode"]} | |
for k, _ in vae_dict.items(): | |
key_name = "first_stage_model." + k | |
full_model['state_dict'][key_name] = copy.deepcopy(vae_model["state_dict"][k]) | |
# Save model with new VAE | |
torch.save(full_model, new_model_name) |
I ran into:
KeyError: 'state_dict'
any idea how to solve it?
@zhuofengli What model are you using? This script will only work on 1.x models.
This error is caused by some merged models, which have been flattened to only the 'state_dict' component already.
Editing the script to this fixes the issues by first checking if the model contains 'state_dict' or is already flattened.
# Check for flattened (merged) models
if 'state_dict' in full_model:
full_model = full_model["state_dict"]
if 'state_dict' in vae_model:
vae_model = vae_model["state_dict"]
# Replace VAE in model file with new VAE
vae_dict = {k: v for k, v in vae_model.items() if k[0:4] not in ["loss", "mode"]}
for k, _ in vae_dict.items():
key_name = "first_stage_model." + k
full_model[key_name] = copy.deepcopy(vae_model[k])
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
oh, this is super useful, thanks! i had no idea vae were stuffed into some of the .ckpt files.