The applicable scope of JiT #48
Description
Does the training of JiT relate to the amount of data? I trained on an infrared dataset with 1,000 images, replacing label embeddings with image feature embeddings in an attempt to reconstruct the original images. However, I found that the embedded features could not be successfully mapped back to the corresponding images, and the generated images lacked clear semantic information. Moreover, there were noticeable inconsistencies between different patches.
