Did you ever try with distilling the VGGT knowledge to the visual tokens just after ViT, instead of the LLM final visual hidden states?

[TyroneLi]

Hi, Did you ever try with distilling the VGGT knowledge to the visual tokens just after ViT saying visual encoder siglip, instead of the LLM final visual hidden states saying inside LLM layers?

thx

[OliverHxh]

Hi, we did not try that, because ViT encoder lacks temporal modeling, which is unable to model the relations of multi-view images.

[TyroneLi]

Hi, we did not try that, because ViT encoder lacks temporal modeling, which is unable to model the relations of multi-view images.

That sounds reasonable, just being curious about this. thanks