Null token embedding missing when using pre-computed text embeddings

jy0205 / Pyramid-Flow

Code of Pyramidal Flow Matching for Efficient Video Generative Modeling

https://pyramid-flow.github.io/

MIT License

2.43k stars 238 forks source link

Closed rob-hen closed 2 weeks ago

rob-hen commented 3 weeks ago

Hi all,

when training with pre-computed text, we need the null token embedding self.null_text_embeds here.

Where do we get this from?

jy0205 commented 2 weeks ago

This is the pre-extracted textual embeddings of empty str, e.g., "", which is used for training with classifier-free guidance.