Gemini: Model is Overloaded

[jch2103]

I testing Gemini 2.o Flash Lite, I frequently (not always) get an error message:

Code Select Expand

04.30 14:26:01+    0 [1E47AC] 02  I> PTCAIConnectorGemini: [oid: 149234] 'D:\Pictures\Temp\Flatirons awards\FileName.jpg' in 500 ms. 0 kw., 0 dlen.
04.30 14:26:01+    0 [15C41C] 02  I> AutoTagger: AI time: 500 ms. 0 output tokens, max_tokens: 229.
04.30 14:26:01+    0 [15C41C] 01  W> AutoTagger: Aborting because of error [600] '[{
  "error": {
    "code": 503,
    "message": "The model is overloaded. Please try again later.",
    "status": "UNAVAILABLE"
  }
}
]'  'V:\develop\IMatch5\src\IMEngine\IMEngineAIAutoTagger.cpp(370)'
04.30 14:26:01+  16 [1E5714] 01  W> UpdateQueue (AutoTagger): Service error 600 '[{
  "error": {
    "code": 503,
    "message": "The model is overloaded. Please try again later.",
    "status": "UNAVAILABLE"
  }
}
]' for file [149234]  'V:\develop\IMatch5\src\IMEngine\IMEngineUpdateQueueAutoTagger.cpp(233)'
I've set up a Gemini API key, but don't know if I left out some other step in the process, or if this is to be expected with the 2.0 Flash Lite model. 

[Mario]

Looks like Google cannot handle the demands. Nothing you or IMatch can do. Try again in a couple of hours.

[jch2103]

I should have first checked the API billing pages: https://aistudio.google.com/app/plan_information and https://ai.google.dev/gemini-api/docs/billing

I've signed up with the Gemini 2.0 Flash 'Free of Charge' plan, which of course has limits. And waiting did allow the request to go through w/o issues.

Still need to test against the other possible models. 
Attached is a spreadsheet comparison of AI results of a few models for two images. The Gemma 3 12B local model did pretty well for these images. 

[Mario]

The attached ZIP file is less than 1KB?
If you want to share your findings, maybe make a single image that shows your test images and the AI results (screen shot of a IMatch File Window is perfect) and post your results and the prompt here: https://www.photools.com/community/index.php/board,149.0.html

[Stenis]

I should have first checked the API billing pages: https://aistudio.google.com/app/plan_information and https://ai.google.dev/gemini-api/docs/billing

I've signed up with the Gemini 2.0 Flash 'Free of Charge' plan, which of course has limits. And waiting did allow the request to go through w/o issues.

Still need to test against the other possible models.
Attached is a spreadsheet comparison of AI results of a few models for two images. The Gemma 3 12B local model did pretty well for these images.

I can say that these bottle necks seems to vanish totally when you accept to pay. The free model comes with extra limitations that are not there when you have activated the Gemini 2.0 Flash.

Wtiting text to 100 pictures doesn´t take more than a minute. Compared to OpenAI Mini (4.1) it is really fast as a flash with Gemini 2.0 Flash Mini. I just tested with 100 pictures and it ran through them in a minute BUT it did miss to write text to 3 or 4 of them. It was easily corrected though in all of them but one. There I had to use the bigger model to solve that problem.

I think it is very good at interpreting texts and nail both species of flowers, trees and places so after all I consider this Gemini 2.0 Flash Mini to be a very substantial improvement in efficiency och productivity compared to the AI-mini-models of the earlier generations we have had before. I think the quality of the texts are very good too. I have reduced the "creativity" to 30% too. Not sure what the impact of that really is but the text might have been a bit more condensed. I will experiment a little more with that, I think.

[Stenis]

The last days I have handled around 1000 pictures from Greece and added texts, keywords and position data. I have used 796 API Calls and used 1302 162 Tokens and it has so far costed 0,49 SEK which is about 5 US cents, so it is very reasonable.

It is really fast even with the bigger Gemini 2.0 Flash model.

The problems I saw initially I have no more.
Today I experienced a problem I haven´t seen before between my prompting for keywords which I earlier had at the keyword prompt but now as there is a slider where you can set how many keywords you want, I have deleted those instructions from the keyword prompt.

The problem I got before was that I suddenly just got one single keyword and then I found the slider and corrected that and adjusted the prompt for the keywords accordingly.

I must say that Gemini Flash Lite is really good at finding places and landmarks in my pictures and that is really helpful and brings me a lot of new memories and info unknown to me as I read those texts. It is pretty exciting and truly remarkable often and really gives me another dimension often to those pictures. A good context is very important for many pictures and makes them much more interesting than without it.

[Mario]

Very well. Let others know how good IMatch and AutoTagger is.

[Stenis]

Very well. Let others know how good IMatch and AutoTagger is.

I already have, but the new models from Google especially has taken the productivity to a whole new level just by the speed they work with.

I guess many photographers from time to time like me have been facing litterally thousands of pictures in post after photo sessions, holiday trips or safaris of different types and if they even use to tag their pictures before adding them to their library they might face a nightmare of metadata maintenance.

With iMatch 2025 and especially Google Gemini 2.0 Flash Lite those nightmares suddenly has turned into something great fun as tasks like that now can be carried out in a few hours instead of days or even weeks.

If I compare the efficency between my old PhotoMechanic workflows and the new AI-driven ones with Autotagger it is really like night and day despite even PhotoMechanic is very fast in some respects but the problem with PM is that that stops at the task of populating Descriptions and Keywords.

Some people have lifted those tasks with CameraBits but from what I know nothing has happened yet to implement AI in PM. It seems like CameraBits is in for a tuff ride because they are facing the need to rewrite the whole foundations of PhotoMechanic since it is still based on old IPTC too.

So we are sitting in a very good place now with iMatch and Gemini thanks to your rapid implementations of these new AI-models the last weeks. Thanks also for  the good, efficient and prompt support-dialogs we have had so far too. They have been extremely important to sort the problems both I and even others had initially. For now I have no problems at all to complain about or get irritated over. It just works and with a speed I never have experienced before!

[Mario]

Maybe you can share some of your findings or the prompts you use over in the AutoTagger Tips & Tricks board.

[Stenis]

Maybe in in a few days.