- Screenshot 2025-04-10 124101.jpg (3.42 MiB) Viewed 34782 times
I know I cover Water Sports as well but its taking it a step to far!
:-) Interesting AI hallucination ;-)
:-) Interesting AI hallucination ;-)
Looks more like a dinosaur too me
I know I cover Water Sports as well but its taking it a step to far!
I know I cover Water Sports as well but its taking it a step to far!
Geoff Mather (G8DHE)
Re: :-) Interesting AI hallucination ;-)
That is actually quite funny...
I wonder what will happen if I throw some lichen images at the AI:
"Orange lichen = Landscape on planet Mars?"
Seriously:
I wonder, can the current implementation at all be used to actually train the AI?
What happens, if you now change the description and image labels telling the AI that, "no, this is no flying crocodile, but actually still a bird". Will it learn???
I am thinking: if I have a bunch of images already tagged in the catalog, will the AI use these tags to describe unused images?
It would be pretty cool thus to be able to train the AI...
I wonder what will happen if I throw some lichen images at the AI:
"Orange lichen = Landscape on planet Mars?"
Seriously:
I wonder, can the current implementation at all be used to actually train the AI?
What happens, if you now change the description and image labels telling the AI that, "no, this is no flying crocodile, but actually still a bird". Will it learn???
I am thinking: if I have a bunch of images already tagged in the catalog, will the AI use these tags to describe unused images?
It would be pretty cool thus to be able to train the AI...
-
andrew.heard
- Posts: 236
- Joined: 15 Jun 10 23:36
Re: :-) Interesting AI hallucination ;-)
I have also found if you Analyse the *same* thumb multiple times, it will return different sets of labels.
Re: :-) Interesting AI hallucination ;-)
I'm not aware that there is a self-training built into these implementations, so I wouldn't expect it to learn, but I maybe wrong!
Biggest problem I can see and it applies to all types and versions is there is no reasoning given behind the choice of a description or it seems is even possible given the nature of the process used - as I understand it?
Have to agree about repeat scans of the same thumbnail, also nearly identical labels created but not quite the same for instance Antenna, Antennae & Antennas all based on the same aerials on the back of my house (being into amateur radio!) none are wrong but why they variations there seems to be some non-deterministic behaviour involved ..... ?
I've over the last few days I've been processing all the images taken from 2024 to date, some 9000+ just to see what use I can make of AI.
So far its been sort of interesting but not really outstanding / useful results, the text descriptions seem OK but not inspiring - more a starting point for submissions to magazines and the like but they become rather bland after a while looking thru them.
Update: No they don't self-learn according to this AI response from a Google of the matter
"AI Overview
No, Ollama AI models do not inherently self-learn from user interactions in the same way that some online AI models might. Ollama is a platform for running and managing local AI models, and the models themselves are typically trained on large datasets beforehand and then potentially fine-tuned for specific tasks. While users can influence the model's behavior through prompts and interactions, the models don't actively adapt or update their internal parameters based on real-time usage, according to a post on LinkedIn. "
Biggest problem I can see and it applies to all types and versions is there is no reasoning given behind the choice of a description or it seems is even possible given the nature of the process used - as I understand it?
Have to agree about repeat scans of the same thumbnail, also nearly identical labels created but not quite the same for instance Antenna, Antennae & Antennas all based on the same aerials on the back of my house (being into amateur radio!) none are wrong but why they variations there seems to be some non-deterministic behaviour involved ..... ?
I've over the last few days I've been processing all the images taken from 2024 to date, some 9000+ just to see what use I can make of AI.
So far its been sort of interesting but not really outstanding / useful results, the text descriptions seem OK but not inspiring - more a starting point for submissions to magazines and the like but they become rather bland after a while looking thru them.
Update: No they don't self-learn according to this AI response from a Google of the matter
"AI Overview
No, Ollama AI models do not inherently self-learn from user interactions in the same way that some online AI models might. Ollama is a platform for running and managing local AI models, and the models themselves are typically trained on large datasets beforehand and then potentially fine-tuned for specific tasks. While users can influence the model's behavior through prompts and interactions, the models don't actively adapt or update their internal parameters based on real-time usage, according to a post on LinkedIn. "
Geoff Mather (G8DHE)
Re: :-) Interesting AI hallucination ;-)
Well, that seems rather unfortunate and not trivial at all.No, Ollama AI models do not inherently self-learn from user interactions in the same way that some online AI models might.
It means you'll have that flying crocodile over and over again, even though you tell the AI that's not what that is...
-
andrew.heard
- Posts: 236
- Joined: 15 Jun 10 23:36
Re: :-) Interesting AI hallucination ;-)
It potentially makes it much more difficult/ unreliable to find anything as there may be multiple labels all describing the same "feature". And if the AI labels are left in a "flat" hierarchy then they are difficult to spot.G8DHE wrote: 11 Apr 25 10:22 none are wrong but why they variations there seems to be some non-deterministic behaviour involved ..... ?
Re: :-) Interesting AI hallucination ;-)
Totally agree! I started just processing a few images too see what sort of results it came up with and I can't say I was overwhelmed by the results. However I decided it really needed a sensible amount so have been processing them month by month going back to the start of 2024 and now have 8300+ images processed and its becoming obvious that I'm not really going to get much benefit it seems.
Its basically processing everything possible as it is unable to understand the context of the image and just what is relevant to identifying the key parts. I mean I really don't need labels such as Green Grass, Blue Sky plus hundreds of other mundane things and even if I was interested say from point of view of a large library of images where a search of "Green Grass" might be relevant it would only be as a function of reducing the volume of subsequent identifications. Also some context could be obtained by linking similar images in a small time frame into a series such as the one above where literally just milliseconds separated the images hence then the switch from Bird to Crocodile could be handled.
I think we have a quite a way to go yet before AI starts to be useful as a general tool. I can see it could be useful even now when the context can be provided and then identifying features in a small subset of such images, perhaps.
Its basically processing everything possible as it is unable to understand the context of the image and just what is relevant to identifying the key parts. I mean I really don't need labels such as Green Grass, Blue Sky plus hundreds of other mundane things and even if I was interested say from point of view of a large library of images where a search of "Green Grass" might be relevant it would only be as a function of reducing the volume of subsequent identifications. Also some context could be obtained by linking similar images in a small time frame into a series such as the one above where literally just milliseconds separated the images hence then the switch from Bird to Crocodile could be handled.
I think we have a quite a way to go yet before AI starts to be useful as a general tool. I can see it could be useful even now when the context can be provided and then identifying features in a small subset of such images, perhaps.
Geoff Mather (G8DHE)
Re: :-) Interesting AI hallucination ;-)
Just realised that there is a painful downside to all these new Labels being created 
When doing a search for label names (top left corner) then now I am getting a dozen or so matches each time as there are so many AI labels, hence making selection of the wanted Label a lot more difficult.
In fact this is making searching by Label name exceedingly difficult, to the point of making this critical facility useless!
We really need an option to turn On/Off any AI generated Labelling across the entire system as a whole as even the use of the Label Panel on the right is becoming very difficult with a zillion Labels appearing when a number of Images are grouped for a listing.
When doing a search for label names (top left corner) then now I am getting a dozen or so matches each time as there are so many AI labels, hence making selection of the wanted Label a lot more difficult.- Screenshot 2025-04-12 115204.jpg (379.06 KiB) Viewed 34651 times
We really need an option to turn On/Off any AI generated Labelling across the entire system as a whole as even the use of the Label Panel on the right is becoming very difficult with a zillion Labels appearing when a number of Images are grouped for a listing.
Geoff Mather (G8DHE)
Re: :-) Interesting AI hallucination ;-)
Now that I'm finally able to use the AI via llama, I'm wondering if the effort was worth the result. If it can't learn from my images then it is not really very useful to label my images. And adding so many pedantic labels does nothing to enable me to filter my images.
Re: :-) Interesting AI hallucination ;-)
The AI using Gemma3 and Ollama isn't very good. Thank goodness I didn't apply the results automatically or I'd be performing a lot of cleanup. I've run the AI on about a hundred images and I bet I applied useable results to maybe 5 or 10 which isn't a great success rate.
Re: :-) Interesting AI hallucination ;-)
After analyzing, don't process results, unless you are a high pace photographer (stock/sports/...), or if the results add value (product photogs for example), or for those where an AI description is more important than no description.
Everything it returns is automatically searchable using the search box and the Advanced Search feature.
Nothing processed = Nothing to cleanup
Already have lots of AI labels from earlier processed results and you don't like these labels then simply delete the AI Labels category. There's no need to keep that if it doesn't add value for you.
Don't expect AI to do the cataloging for you. Expect it to help you. In most of the photos that I have analyzed, there was at least something, somewhere in the results (descriptions) that adds value and offers search terms that I would never use as catalog labels. Still, they automatically add to the search results.
Then again, if mostly what you use for searching are the catalog labels, then why bother at all.
PS. Gemma3 is, afaik, the best model available for offline processing with Ollama. But it still is just a ~4GB model, which is nothing compared to the models from OpenAI or other major AI players. Ollama works offline and is great when (and only when) privacy is an issue.
Everything it returns is automatically searchable using the search box and the Advanced Search feature.
Nothing processed = Nothing to cleanup
Already have lots of AI labels from earlier processed results and you don't like these labels then simply delete the AI Labels category. There's no need to keep that if it doesn't add value for you.
Don't expect AI to do the cataloging for you. Expect it to help you. In most of the photos that I have analyzed, there was at least something, somewhere in the results (descriptions) that adds value and offers search terms that I would never use as catalog labels. Still, they automatically add to the search results.
Then again, if mostly what you use for searching are the catalog labels, then why bother at all.
PS. Gemma3 is, afaik, the best model available for offline processing with Ollama. But it still is just a ~4GB model, which is nothing compared to the models from OpenAI or other major AI players. Ollama works offline and is great when (and only when) privacy is an issue.
This is a user-to-user forum. If you have suggestions, requests or need support then please send a message
Re: :-) Interesting AI hallucination ;-)
The differences between AI Models illustrated for the same image. Model size matters!
Which model can identify this deer (or is it an elk)?
Spoiler alert - it's the largest model: GPT 4.1.
Using Ollama and Google Gemma3:
Using OpenAI and GPT4.1-Nano:
Using OpenAI and GPT4.1-Mini:
Using OpenAI and GPT4.1:
Which model can identify this deer (or is it an elk)?
Spoiler alert - it's the largest model: GPT 4.1.
Using Ollama and Google Gemma3:
- deer-gemma3.png (316.35 KiB) Viewed 34502 times
Using OpenAI and GPT4.1-Nano:
- deer-gpt41-nano.png (330.18 KiB) Viewed 34502 times
Using OpenAI and GPT4.1-Mini:
- deer-gpt41-mini.png (516.58 KiB) Viewed 34502 times
Using OpenAI and GPT4.1:
- deer-gpt41.png (314.36 KiB) Viewed 34502 times
This is a user-to-user forum. If you have suggestions, requests or need support then please send a message
Re: :-) Interesting AI hallucination ;-)
Another one, this time a city photo in NY.
There's only one that was able to identify the building and that's, again, the largest model.
With Google Gemma3:
With OpenAI-41-Nano:
With OpenAI-41-Mini:
With OpenAI-41:
There's only one that was able to identify the building and that's, again, the largest model.
With Google Gemma3:
- ny-gemma3.png (628.72 KiB) Viewed 34491 times
With OpenAI-41-Nano:
- ny-gpt41-nano.png (630.87 KiB) Viewed 34491 times
With OpenAI-41-Mini:
- ny-gpt41-mini.png (624.67 KiB) Viewed 34491 times
With OpenAI-41:
- ny-gpt41.png (632.12 KiB) Viewed 34491 times
This is a user-to-user forum. If you have suggestions, requests or need support then please send a message
Re: :-) Interesting AI hallucination ;-)
This time a landscape photo. I am really stunned that GPT41 is able to guess the exact location of this picture.
With Google Gemma3:
With OpenAI-4.1 Nano:
With OpenAI-4.1 Mini:
With OpenAI-4.1:
With Google Gemma3:
- garda-gemma3.png (801.19 KiB) Viewed 34489 times
With OpenAI-4.1 Nano:
- garda-gpt41-nano.png (812.14 KiB) Viewed 34489 times
With OpenAI-4.1 Mini:
- garda-gpt41-mini.png (807.89 KiB) Viewed 34489 times
With OpenAI-4.1:
- garda-gpt41.png (805.1 KiB) Viewed 34489 times
This is a user-to-user forum. If you have suggestions, requests or need support then please send a message
Re: :-) Interesting AI hallucination ;-)
One last one, and let's make it more difficult. Let's see which model can identify this statue in a non-closeup photo.
This is the statue and I have no idea what it is...AI to the rescue?
Gemma3 thinks it's a person, not a statue:
OpenAI-4.1 Nano sees a rock, not the statue on it.
OpenAI-4.1 Mini knows its stuff, and even informs me who the artist is and when it was installed. Amazing!
OpenAI-4.1 also identifies the statue:
This is the statue and I have no idea what it is...AI to the rescue?
- statue.png (624.98 KiB) Viewed 34486 times
Gemma3 thinks it's a person, not a statue:
- statue-gemma3.png (596.32 KiB) Viewed 34486 times
OpenAI-4.1 Nano sees a rock, not the statue on it.
- statue-gpt41-nano.png (612.16 KiB) Viewed 34486 times
OpenAI-4.1 Mini knows its stuff, and even informs me who the artist is and when it was installed. Amazing!
- statue-gpt41-mini.png (623.09 KiB) Viewed 34486 times
OpenAI-4.1 also identifies the statue:
- statue-gpt41.png (606.31 KiB) Viewed 34486 times
This is a user-to-user forum. If you have suggestions, requests or need support then please send a message