Skip to content

Question about position embeddings #238

@zlenyk

Description

@zlenyk

Hi, thank you for your work and sharing it. I wanted to ask a question somewhat related to 214. I'm looking into fine-tuning the model so that it would keep only last K frames as conditioning. However in current implementation of assigning IDs to images, it seems that the oldest frame gets 0 and newest gets N. Is there a reason it's not the other way? It would seem to me a bit more natural that the embeddings are fixed from the perspective of current generation (the one right before always has 0). I'm wondering if it's worth expeerimenting with fine-tuning this model, with this reversed.
Would you have some thoughts on this?

Thank you

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