metadata
license: mit
library_name: transformers
pipeline_tag: zero-shot-image-classification
You probably do not need this unless you are training your own IP Adapters.
Modified version of the vision encoder of CLIP-ViT-H-14-laion2B-s32B-b79K to handle 448 x 448 inputs vs the original 224 x 224 inputs. It will probbaly not work for classification (as is), but will DIP work for for IP+ adapters that use CLIP-ViT-H, though they will need to be fine tuned a little more.
Hidden layer outputs go from (257, 1280) to (1025, 1280), which can be digested by the Resampler without modification or weight resizing.