Skip to content

DINO feature in sampling process #6

Description

@kowoonho

Thank you for your insightful work!
I have a question regarding the sampling process.

In LDM, the image latent produced by the VAE is regularized via a KL loss, so starting from N(0, I) during sampling works as intended.
However, DINO feature maps used as conditioning are not regularized to follow any prior distribution.

During sampling, without a reference image, should we also initialize the DINO features from N(0, I)?
Wouldn’t this lead to a distribution mismatch, since DINO features are not trained to follow a Gaussian prior?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions