You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, thanks for your brilliant work! I've tried diffusion classifier with stable dffusion 2.1 locally, and found the performance on cifar10 achieves 85.88 which is even higher than illustrated in the paper. This is a proof of the power of discriminative ability of generative diffusion models.
My question is, have you tried other diffusion models, like DDIM, unCLIP and Score SDE? Is the impressive performance highly dependent on stable diffusion?
The text was updated successfully, but these errors were encountered:
I am not the author. Based on my understanding, the authors propose an algorithm that leverages the discriminative ability of existing diffusion models for classification tasks. That means the performance is determined by the algorithm used and the diffusion models. And I personally think diffusion models are more important to the performance improvements. Nonetheless, developing effective algorithm is still valuable. That is why I ask the author to compare with exact log-likelihood, another alg, see #2
@pengzhangzhi Hi, thanks for your reply! I've just checked #2 and the Score SDE paper, and found your observation quite inspiring. It seems there's some inherent connection between probability flow ODE and class discrimination. As the auther stated in #2 , how to effectively and efficiently estimate log-likelihood could be a problem to be solved next.
Hi, thanks for your brilliant work! I've tried diffusion classifier with stable dffusion 2.1 locally, and found the performance on cifar10 achieves 85.88 which is even higher than illustrated in the paper. This is a proof of the power of discriminative ability of generative diffusion models.
My question is, have you tried other diffusion models, like DDIM, unCLIP and Score SDE? Is the impressive performance highly dependent on stable diffusion?
The text was updated successfully, but these errors were encountered: