Skip to content

Weight initialization for training ControlNet #39

@kietngt00

Description

@kietngt00

Thank you for sharing the great work!

I would like to ask how the weight of the transformer of ControlNet can be initialized from the Stable Diffusion 3.5 model for training?
Particularly, when I check ControlNet for Canny and Depth, the transformer uses SD3SingleTransformerBlock, while SD3.5's transformer uses JointTransformerBlock.

Is it correct that ControlNet's transformer just copies the query, key, value, output projection weights, and the FFN from the SD3.5 transformer?

I am looking forward to receiving your answer.
Much appreciate!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions