Pixel Decoder

A decoding module that restores the semantic hidden states of a vision-language model to the image grid and upsamples them to the original image resolution, supporting dense prediction tasks such as segmentation and matting.