single source · model architecture
Feature-wise Linear Modulation (FiLM)
Feature-wise Linear Modulation (FiLM) conditions a network's intermediate activations on another input, here language embeddings.1 It scales and shifts each hidden dimension of the visual encoder's embeddings, with learned per-block projectors, so language modulates visual features.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
The functions and that project language embeddings to obtain and are implemented as simple affine transformations
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successentire feature maps are scaled and shifted by individual elements of and
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success
Also called FiLM.
Filed under model architecture › channel mixer