This desperately needs a text search feature!

The first thing I thought of was 'oooh, maybe put a pic on the landing page of this website I help out with' but I can't search for e.g. 'chess' or whatever theme I want.

The whole point of the site is to annotate the photos, but then to not make them searchable by annotation? Shame!

Agreed and it's nice to hear it's first on the todo list.

Until then, though, just change the word in the tag (with hyphens for multiple words), e.g.:

https://taken.photos/tags/chess/ https://taken.photos/tags/tabletop-game/

That's the first thing in my todo list. Thanks for clarifying.

Where do the images come from? The "About Us" page says they're all public domain, but this image has a copyright: https://taken.photos/photos/pink-flower-on-plant-3753020/

My favourite is "A statue of a cement bench"

Interesting. You _really_ need some way for folks to suggest improved captions.

A tree with a mountain in the background.... completely ignoring the house which is the central focus of the photo.

https://taken.photos/photos/tree-with-mountain-in-background...

or the aforementioned "A cat sitting on top of each other"

I'm thinking this is a great _starting point_ a way to bulk ingest lots of photos, but as a resource i'm going to loose faith in it every time i see an obvious AI screw up. Technically yes, they're still useful titles that help (well, they would help if you had a search which I can only assume you're planning because otherwise what's the point) but _emotionally_ i'm going to be "yeah but they're always wrong" (even if it's not true)

How about "A man and a dog walking on a sandy beach?" https://taken.photos/photos/man-dog-walking-on-sandy-beach-3...

This is the state of modern AI. Mostly it's still utter garbage.

What do you use for the AI labelling? Looks pretty accurate, but fun to see the random mistakes

A zebra standing on top of a building https://taken.photos/photos/zebra-standing-on-top-of-buildin...

"A cat sitting on top of eachother" https://taken.photos/photos/cat-sitting-on-top-of-each-other...

some of these are comedy gold :)

Nice work. But I would suggest to add a feature to report wrong tags. Perhaps that could even lead to optimize the classification AI or you could at least review the suggestions and delete wrong tags manually?

Suggestion when adding text search: allow “and” tags, such as “outdoor AND person” or “people AND group AND smile”

Is there anyway to read more about the technical bits regarding how this is done, i.e. how are the annotations made?

Sure, why not. I scrape multiple Russian websites with photos and annotate pics with https://azure.microsoft.com/en-us/services/cognitive-service...

Surprisingly MS does much better than Google and their https://cloud.google.com/vision/

I tried both and got frustrated with almost random results from Google.

Those platforms you're sourcing the photos from presumably have content restrictions, which causes the "nude" tag to be dominated by photos without any humans in them. For example this flame: https://taken.photos/photos/2568745/

Well, there was a bug in my data migration/cleanup script for tags. Thanks for catching.

The 'curated collection' statement means 'there is no adult or NSFW' content. AI does great job with detecting nudity, and as a pragmatic person I decided to keep all scraped adult content and have it served at https://nudes.best/