Skip to content

Question about setting eta=1.0 in DDIM — does it reduce to DDPM? Could lower values like 0.5 be used? #24

Description

@hugluckylu

Hi,authors,
Thanks for the great work and for open-sourcing the code!
I have a question regarding the choice of eta=1.0 in the DDIM sampling configuration. As far as I understand, setting eta=1.0 makes DDIM equivalent to DDPM, which reintroduces full stochasticity. I’m curious about the motivation behind this — was it intentionally done to introduce randomness, or is there another reason?
Also, I wonder if using a lower value (e.g., eta=0.5) would be valid in this context, as it might offer a trade-off between sample quality and diversity. Have you experimented with such values?
Looking forward to your insights. Thanks again!

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