Collected sources and patterns will appear here. Add from search or the patterns library.
Tensor<VisualFeatures> -> Tensor<TextEmbeddings>
Align visual token features with the language model's hidden representation space using a Multi-Layer Perceptron.
Problem it solves
Visual tokens and text embeddings operate in different dimensional spaces, preventing direct interleaving in decoder attention blocks.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.