Here's the scenario I've been able to reproduce:
I've got an upload form using a table or collection to set a few metadata fields. It has an upload button, and I've set the button to show a medium thumbnail of the files uploaded. If enough photos are uploaded to the form (over 60-100), then when the save to repository step goes off, only about 5-6 pages will include a thumbnail; after about 50 images (out of 200-ish), the rest of the pages just list attachment names. When this happens to the upload report, one of the images randomly gets saved to the repository with an extra 20,000-30,000 pages of text that look like a photo encoded as text. The save to repository step will also take about 10-20 times longer than it should to save (no doubt chugging to output the report with thumbnails).
I've included a sample of the output of the report. Note each file is a photo that should have a thumbnail, but they stop after about 60 thumbnails. There's another thumbnail on page 9, but there's no rhyme or reason to it. Each of the photos is between 5 and 13 megabytes; for a total of about 1.2GB in this sample.
For now, I've addressed this by turning off the thumbnails for the upload report; the save to repository step goes through in about 2 minutes without corrupting any of the images.