Is Animatediff used for inference , or just training only?

[moonsunstars39]

Hello,

I tried to run the inference demo without Animatediff model, it seems like running ok. Is the AnimateDiff motion module necessary for the inference process? 'cause I couldn't find anywhere import the motion module in your code except training process.

In addtion, can I use Animatediff v3 instead of v2 ? I have this odd question because you placed mm_v2.ckpt in the inference example :D

Thank you in advance.

[theFoxofSky]

It is just for training. It remains in the inference code because the inference code is modified based on the training one.

[moonsunstars39]

It is just for training. It remains in the inference code because the inference code is modified based on the training one.

Thank you :D

[Russell-Zeng]

In the "3.1. Implementation Details" section, it is mentioned that "In motion learning, only motion modules, which are initialized from AnimateDiff [8], are optimizable." I would like to understand why the motion module trained in the second phase doesn't participate in inference. Could you clarify the reason behind this decision?
many thanks

[theFoxofSky]

In the "3.1. Implementation Details" section, it is mentioned that "In motion learning, only motion modules, which are initialized from AnimateDiff [8], are optimizable." I would like to understand why the motion module trained in the second phase doesn't participate in inference. Could you clarify the reason behind this decision? many thanks

Here, the not-used one is the initialization parameter, the trained one is used