Clarification on C style Generation: Difference Between StyleNet and StylePyramidNet

[theohsiung]

Hi,

Could you please explain the process of C-style generation during inference? I noticed that in the pre-training phase, the system uses StyleNet, whereas during inference, it switches to StylePyramidNet.

Could you clarify the differences between these two architectures and their roles in the pipeline?

Thank you very much for your help!

[weichaozeng]

Sure. These two architectures share a same VisionTransformerEncoder module as backbone for capturing style features, while the StyleNet contains additional GlyphHead and SpatialHead for multi-tasks pretraining and the StylePyramidNet contains an additional pyramid module for multi-scale features extraction which is used in both training and inference.