Why is the temporal layer trained inStage 2 not initialized using the temporal layer weights in AnimateDiff?

[chlinfeng1997]

Thanks for open source！ The original paper mentions using the temporal layer weight of AnimateDiff for initialization, why does it not appear in the code? Looking forward to your reply！

[guoqincode]

Hi, you can combine unet, which you trained in the first stage, with ckpt from animatediff.

[guoqincode]

just like:

import torch

def merge_ckpts(ckpt1_path, ckpt2_path, output_path): ckpt1 = torch.load(ckpt1_path, map_location="cpu") ckpt2 = torch.load(ckpt2_path, map_location="cpu") merged_state_dict = {ckpt1, ckpt2} torch.save(merged_state_dict, output_path)

trained_unet_path = './pretrained_models/unet_stage_1.ckpt' motion_module_v1_path = './pretrained_models/AnimateDiff/mm_sd_v15.ckpt' merged_ckpt_path = "./pretrained_models/merge_trained_unet_motion.ckpt"

[L1129433134]

Why don't you use it? Is the performance not good?！ @guoqincode

[guoqincode]

Why don't you use it? Is the performance not good?！ @guoqincode

I used it.

[zhanchao019]

Why don't you use it? Is the performance not good?！ @guoqincode

I used it.

Have you do comparison experiments? Is training temporal module from scratch much different from fine-tuning networks?