Skip to content

Question about Distillation Details #34

@UknowSth

Description

@UknowSth

Thank you for releasing such an excellent open-source work. I would appreciate some clarification regarding the distillation procedure described in the paper.

As mentioned, the Teacher Model provides supervision for the generation of each block (e.g., 5 seconds). Since the Student Model conditions on previously generated frames as historical context, it is not entirely clear to me whether the Teacher Model also operates under the same setting.

Could you please clarify whether the Teacher Model performs text-to-video denoising directly, or whether the teacher model (e.g., Wan2.1-14B) is first adapted to incorporate historical frame conditioning before being used for supervision?

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