A new preprint reports favorable classification, detection and segmentation results from adapter modules attached to a frozen image-compression system. The approach uses far fewer trainable parameters than full fine-tuning, while the study remains limited to its named benchmark settings.
