I’ve done a data export of my ChatGPT data a few times, and observed this (file size, filename):
836M chatgpt-2025-03-10.zip
304M chatgpt-2025-03-30-02-24-17.zip
465M chatgpt-2025-03-30-19-44-23.zip
Now the later ones generally should have a few more generated images and a bit more chats, so I would have assumed them to be larger. Especially the size difference between the two exports from the same day are suspiciously large.
Looking at the dalle-generations/ directories, the first one has 1897 images, the second one 540, and the third one 1059.
The images are still referenced in the asset_pointers in conversations.json and visible in the ChatGPT user interface if I go to the relevant session, but they’re (seemingly randomly) absent from the dalle-generations/ directory. My conversations.json contains 2746 unique file-service:// URIs, which I think are all dall-e generated images.