Issues with x_t (noise) prediction and some question about the training loss.

[kaipengfang]

1. Based on Issues Loss weights used during training #19 and the config, Geometric losses are not used. Instead, they are implemented through the 263-dimensional motion representation. Therefore, geometric losses will never be used; they are just redundant code. Is my understanding correct?
2. Why use x_0 as the training target? Is there an explanation for this? Currently, most mainstream diffusion models predict xt instead.
3. When I set the code to predict x_t as the training target, I found that the final results were very poor (
   Training for 23,000 steps, use linear beta scheduler, the character keeps shaking and doesn't perform the action well). What could be the reason for this? I only made the change to set self.model_mean_type = ModelMeanType.EPSILON in the training and sampling code. Could such a simple modification cause any issues?

predict x_t sample results：

sample01.mp4

sample02.mp4

[GuyTevet]

Try to also change the noise scheduling to linear instead of cosine.

[kaipengfang]

In fact, I have tried and the problem still exists.

…

---- Replied Message ---- | From | Guy ***@***.***> | | Date | 01/30/2025 15:55 | | To | GuyTevet/motion-diffusion-model ***@***.***> | | Cc | BYTEGATHER ***@***.***>, Author ***@***.***> | | Subject | Re: [GuyTevet/motion-diffusion-model] Issues with x_t (noise) prediction and some question about the training loss. (Issue #232) | Try to also change the noise scheduling to linear instead of cosine. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>