My understanding is that an image is by itself PII, regardless of whether or not it has any additional information associated with it. I don’t think there’s a way to retain images without contravening GDPR.
When looking at a single datum by itself, this seems to rule out anything except PII i.e. data that identifies or can be used to identify an individual.
What that says is that, if (A,B,C) identifies a person, each of A, B, and C, in isolation, is personal data, not that you will be allowed to keep the pair (A,B) if it doesn’t.
One mathematically can cut each bit of information in units of arbitrarily small entropy. So, if taken to the letter, “this user is not Mark Zuckerberg” would be personal data. I doubt jurisprudence will go that far, but we’ll see.
Whether information that can only be used to identify someone but doesn't tell you anything useful about them is still personal data is unclear to me.
Not in this case, because if the photos or videos contain recognisable people then they are themselves personal data.
How far the new subjects rights involving data deletion will go in practice is one of the biggest unknowns with the GDPR. Clearly from a technical point of view we understand that deleting a key isn't the same as deleting data from a disk, and often that would also include deleting a file in a filesystem if the underlying storage isn't robustly wiped as well. Throw in the kinds of distributed architecture, redundancies and backup systems that many organisations use, particularly in the era of cloud-based hosting and off-site backup services, and you have an unfortunate conflict between not truly deleting data (and therefore still having some degree of risk that the data will leak even if it's intended to be beyond use, contrary to the spirit and possibly the letter of the new regulations) and potentially high or even prohibitive implementation costs to ensure robust deletion of all copies of personal data when a suitable request is received.