Jump to content

Maybe - Coming soon to Alamy?


Recommended Posts

As a contributer to some of the POD sites, I was amazed to see one of them have introduced A.I. descriptions and keywords added automatically by default when you upload an image. They describe it as Automated Image Keywords And Descriptions, and it works really well. You have to tweak the descriptions as it only describes the image without adding where it is and the exact description or names of buildings, but it adds 10 really good keywords which I hadn't always thought of. It certainly saves a lot of time.

The sign of things to come?

  • Like 2
Link to comment
Share on other sites

Since the POD site I contribute to has added AI captioning every mountain has become ‘majestic’…. and the captions are full of superfluous romanticised fluff.  I’m not in the least bit impressed.

 

Mark

  • Like 2
Link to comment
Share on other sites

3 hours ago, geogphotos said:

Alamy has zero hits for 'unmajestic mountain'🫣

Lets try this...

9,168,378 on Alamy for mountain NOT majestic

317,946 on Alamy for mountain AND majestic

So, on Alamy only 3.5% of Mountains are majestic

 

But, upload an image of a mountain scene to this particular POD site and its AI captioning will usually add the term majestic... What defines a "majestic mountain" I wonder? Any hill that's wider at the bottom than the top?

 

Mark

 

 

Edited by M.Chapman
  • Like 2
Link to comment
Share on other sites

20 minutes ago, Mr Standfast said:

So why is AI denoise so good and AI keywording so rubbish?

 

There's a Phd in there somewhere, off you go!

 

G'night. 😉

 

 

IMO - AI image keywording has been trained by “harvesting” a large sample of the keyworded images out there already. Unfortunately, this training data is “contaminated” by the high proportion of badly keyworded images there are out there. Whereas AI noise reduction has been trained on pairs of images of the same scene with and without noise (e.g. pairs of images of identical scene taken at low and high ISO settings, or clean images with and without simulated noise added) so the training data is much more robust/reliable. That’s my theory anyway. For AI to work properly it’s essential to have validated training data, otherwise it’s garbage in garbage out. There’s too much misplaced confidence in AI as if it’s capable of solving any problem. Sure it can work really really well, but there are limits.

 

Mark

  • Love 1
Link to comment
Share on other sites

3 hours ago, Avpics said:
Search Term UCO Sales
Descending
Zooms Views CTR (%)
Majestic Mountain 1 0 0 200 0.00

 

That's one UCO, 200 views and zero zooms in the last year on AoA. Zero CTR

 

Probably this client was searching for the real Majestic Mountain.

 

I'm not that concerned about AI keywords, but much more about AI searches by clients.

 

wim

  • Like 1
Link to comment
Share on other sites

I've used ChatGPT 4 to suggest keywords and it is useful in coming up with things I hadn't thought of, but it can be hilariously wrong, too. For instance, it thought these were actual security vehicles and suggested things like "surveillance", "retail safety", "modern security", and "electric vehicles". 

 

"These keywords reflect the objects in the image (patrol vehicles, electric carts), their purpose (mall security, surveillance), the setting (shopping mall, mall interior), and related themes of safety and property management (retail safety, property protection). They could help the image be discovered by users interested in security services, retail management, or commercial real estate."
 

mall-patrol-smarte-carte-strollers-lined

Edited by Mark Scheuern
Link to comment
Share on other sites

17 hours ago, Mr Standfast said:

So why is AI denoise so good and AI keywording so rubbish?

 

There's a Phd in there somewhere, off you go!

 

G'night. 😉

 

 

Programmers who knew photography and image processing wrote AI Denoise.   People market AI as a keywording solution without understanding the tenuousness of the link between images and language.  People who didn't understand what people looked for wrote keywording software for a quick buck.   People who are first rate graphics application programmers are going to work for a handful of companies that pay serious money for serious results.

Link to comment
Share on other sites

22 hours ago, Mr Standfast said:

So why is AI denoise so good and AI keywording so rubbish?

 

 

 

 

One is a more difficult problem than the other. AI denoise is really good. AI keywording is going to get better very quickly. The number of things we humans do better is going to rapidly diminish. 

Link to comment
Share on other sites

17 hours ago, Mark Scheuern said:

 

One is a more difficult problem than the other. AI denoise is really good. AI keywording is going to get better very quickly. The number of things we humans do better is going to rapidly diminish. 

Unfortunately, with respect to keywording I fear AI is heading in the opposite direction. Images tagged by AI are progressively contaminating the data sets used to train new AI algorithms.

 

Mark

  • Love 1
Link to comment
Share on other sites

That's an interesting problem, for sure: AI training on AI-generated data. I wonder how they're going to try to deal with that? It's easy to imagine a good deal of the content of the Internet being AI-generated in the coming years. 

Link to comment
Share on other sites

 

On 22/03/2024 at 18:16, Paul Thompson said:

The sign of things to come?

 

Yes! I have been impressed enough by the results that I made an browser extension to help me go through the keywording faster. I am happy with the great majority of keywords proposed. I never got the "majestic" one, but often get the "tranquil" one. 

 

 

On 23/03/2024 at 15:09, Mark Scheuern said:

I've used ChatGPT 4 to suggest keywords and it is useful in coming up with things I hadn't thought of, but it can be hilariously wrong, too. For instance, it thought these were actual security vehicles and suggested things like "surveillance", "retail safety", "modern security", and "electric vehicles". 

 

"These keywords reflect the objects in the image (patrol vehicles, electric carts), their purpose (mall security, surveillance), the setting (shopping mall, mall interior), and related themes of safety and property management (retail safety, property protection). They could help the image be discovered by users interested in security services, retail management, or commercial real estate."
 

 

 

 

I have been using GPT4's Vision API as well. For that image, I get "Row of blue mall patrol carts parked inside an empty shopping center with tiled flooring and a closed store in the background", which is not accurate indeed, but also I think a better starting point than blank.

 

For example, I have an image of a guanaco. Close enough for me to recognize it's a guanaco. But giving it to GPT without any additional information, I was given the keywords "guanaco", "lama guanicoe", "South America", "ungulate", "mammal", "wildlife", "conservation" ... I could have come up with those myself, but definitely not as fast as with GPT assistance.

Link to comment
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
×
×
  • Create New...

Important Information

We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.