Skip to content

details about the training of zimage_turbo_training_adapter #1486

@LPengYang

Description

@LPengYang

Thank you for your tremendous contributions to the community. I have a question regarding the de-distillation LoRA strategy described in the repository. To enable fine-tuning of the few-step distilled model (zimage-turbo), the approach first generates a synthetic dataset using the distilled model itself, and then trains a de-distillation LoRA on this data. My specific question is: when training this de-distillation LoRA, is the optimization objective simply the standard flow matching loss (i.e., corrupting clean samples with noise to obtain noisy states and predicting the velocity), or does it employ distillation-specific losses?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions