[Question] Can you explain the tensor shapes for EmptyLatentImageLarge?

[vedantroy]

The AnimateDiff paper states that inputs take the form: B x C x F x H x W and are reshaped to (B x F) x (C x W x H).
I was reading through this code-base to understand the implementation, and I see:

ComfyUI-AnimateDiff-Evolved/animatediff/nodes.py

Line 437 in 29dee14

latent = torch.zeros([batch_size, 4, height // 8, width // 8])

Am I correct in assuming that here:

• B = 1 (batch size 1)
• F = batch_size
• C = 4 (the number of input channels, presumably after encoding using the VAE?)
• W/H are self explanatory

[vedantroy]

And a quick follow-up question as well, what is the best way to re-use the AnimateDiff beta schedule / noising schedule in a custom node? I'm trying to write a custom node which noises the latents using the AnimateDiff schedule.

So far, I have a basic node that accepts an encoded image, repeats it frames number of times. But I'm stuck on the part where you apply the correct forward schedule to the latents.

from einops import repeat

class Example:
    def __init__(self):
        pass
    
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "latent": ("LATENT",),
                "frames": ("INT", {
                    "display": "number"
                }),
            },
        }

    RETURN_TYPES = ("LATENT",)
    FUNCTION = "generate"
    CATEGORY = "Cantaloupe"

    def generate(self, latent, frames):
        anim_lat = latent["samples"].clone()
        anim_lat = repeat(anim_lat, 'b c h w -> (b frames) c h w', frames=frames)

        # TODO: What goes here?
       
        return ({"samples": anim_lat},)