Questions about the predicted_noise of EDM framework.

pixeli99 / SVD_Xtend

Stable Video Diffusion Training Code and Extensions.

574 stars 55 forks source link

Questions about the predicted_noise of EDM framework. #19

Closed Pandint closed 3 months ago

Pandint commented 8 months ago

Sorry to disturb you again.

I am trying to accelerate SVD with DDIM, but in the EDM training framework, I didn't find how to calculate the predicted_noise with sigmas and model_pred. Can you help me?

Pandint commented 8 months ago

I derivate from the EulerDiscreteScheduler and obtain the results as follows: denoised_latents = model_pred * (-sigma / (sigma**2 + 1) ** 0.5) + (noisy_latents / (sigma**2 + 1)) predicted_noise = model_pred * (1 / (sigmas**2 + 1) ** 0.5) - (noisy_latents * sigmas/ (sigmas**2 + 1)) I think this might be right.

pixeli99 commented 8 months ago

Hi, actually I don't quite understand your motivation. Are you looking to change the sampling method of SVD to DDIM?

Pandint commented 8 months ago

Yes. And DDIM needs to take pred_noise and model_pred as input.