I always got noisy images!

[109km]

I'm using MacOS M2 chip, when I'm working with other models, everything works fine.
And when using the motion model the output image is always noisy or got weird image.

What's wrong with the workflow? Or some other problems?

workflow.json

[Kosinkadink]

The batch size determines the total animation length, and in your workflow, that is set to 1. The amount of latents passed into AD at once has an effect on the actual output, and the sweetspot for AnimateDiff is around 16 frames at a time. Since you are passing only 1 latent into the KSampler, it only outputs 1 frame, and it is also very deep fried.

If you increase the amount of latents to ~16, you will get expected outputs. If you have a Context Options node attached, if your passed-in latents (in your workflow, that is determined by the batch_size on your Empty Latent Image node) exceed the context_length, then it will begin to use a sliding context window to sample context_length latents at a time to allow you to produce longer than sweetspot amount of frames at the sweetspot quality.

Once you increase your batch size as AD requires, you will almost certainly run into the pytorch bug that exists for M1 and M2 macs. The workaround is mentioned here: #48 (comment)

It is also recommended to use the --use-split-cross-attention startup argument on Mac to avoid other issues/crashes that could pop up.

Let me know if that helps!

[Kosinkadink]

Soon(ish), I'll make it easier to implement that workaround on Macs.

[109km]

It works!
Thank you so much!