I used to be able to prompt ChatGPT to describe an image based on a link to an image I uploaded to an image-sharing site (imgur.) and it would provide me with a response - sometimes accurate, sometimes not so much. Now, ChatGPT indicates that it cannot describe an image based on a link. Did something perhaps change?
It never had any such ability. The multimodal model was never publicly released.
Ohh… Was that possible? I never tried it then
It was not ever possible for the general public.
GPT has never been able to browse the internet. It was also really bad at telling you it couldn’t, and would instead make things up based on Chat context or the keywords in the URL.
What you’re seeing now is likely a shift in behavior to more accurately convey it can’t access links.
It cannot do that as stated by folks here with the model we have, I tried this at the very first time of the release with plugins, if the name of the image/link contain a hint it might can just guess based on that.
Other than that, whatever it used to say about the image links I gave was pure clever hallucination.
The way to make it to describe any image with the current model is to have a plugin that can call a specialized model in hugging face for example, this is the only way to get the right answer or for anything else beyond the capabilities of the current model we have.
It can’t. Probably it watched some keywords in the url, for example if in the url there was “beautiful_cat” in a point, it deduced that it was an image of a cat
Ah … this is it! It was responding to text in the URL. Thank you! One of the challenges with ChatGPT and other black box AI is that it’s difficult to determine the full capabilities of such tools.
It turns out I was mistaken, as another fellow point out in a later reply. ChatGPT was responding to the file name, and not the image itself.
A possible explanation for ChatGPT’s description of an image url is that it was based on the subdirectory paths of the url. For example, website.extension/motorbike/harley/black/500x500.jpg?affdsfsfdsfsdgrehge might suggest an image of a black Harley motorbike.