Clarification on DINOv2 Distillation in Video DC-AE Checkpoint
#2
by
jmkim0309
- opened
Hello,
In the technical report, it is mentioned that the video DC-AE model is further tuned using DINOv2 distillation loss after the initial training.
However, I noticed that the distillation step is not included in the training code available on the OpenSora GitHub repository.
This leads me to some confusion — is the released checkpoint (F32T4C128_AE.safetensors) the distilled version of the model, or not?
Could you please clarify this?
Thank you in advance!