Vision Tower
The subnetwork in a multimodal model responsible for encoding images into visual tokens, typically based on a ViT-style architecture. Its output is aligned to the language model's embedding space through a projection layer, and compact designs reduce computation by lowering resolution or reducing the number of layers.
Timeline (last 90 days)
Change #19247 is titled 'preserve small-image detail for compact Gemma 4 vision towers', aiming to preserve detail information in small-sized images when Gemma 4's compact vision encoders process images.
A vision tower is the subnetwork in a multimodal model responsible for encoding images into visual tokens, usually stacked from ViT-like structures, with outputs aligned to the language model's embedding space through a projection layer.
All Facts (2)
Change #19247 is titled 'preserve small-image detail for compact Gemma 4 vision towers', aiming to preserve detail information in small-sized images when Gemma 4's compact vision encoders process images.
50%UnverifiedA vision tower is the subnetwork in a multimodal model responsible for encoding images into visual tokens, usually stacked from ViT-like structures, with outputs aligned to the language model's embedding space through a projection layer.
50%