Using gpt-4o vision to understand images

OpenAI released gpt-4o recently, which is the new flagship model that can reason across audio, vision, and text in real time. It's a single model which can be provided with multiple types of input (multi modal) and it can understand and respond based on all of them. 

The model is also available on Azure OpenAI and today we are going to have a look at how to work with images using the vision capabilities of gpt-4o. We will be providing it with images directly as part of the chat and asking it to analyse the images before responding. Let's see how it works:

We will be using the Azure OpenAI service for working with the OpenAI gpt-4o and since we will be using .NET code, we will need the Azure OpenAI .NET SDK v2:

1. Basic image analysis

First, let's start with a simple scenario of sending an image to the model and asking it to describe it.

2. Answer questions based on details in images

Next, let's give a slightly more complex image of  some ingredients and ask it to create a recipe:

Image source: allrecipes.com

3. Compare images

This one is my favourite, let's give it 2 images and ask it to compare them against each other. This can be useful in scenarios where there is a single "standard" image and we need to determine if another image adheres to the standard.

4. Binary data

If the URL of the image is not accessible anonymously, then we can also give the model binary data of the image:

5. Data URI

We can also use Data URI's instead of direct URLs

6. Limitations

As per OpenAI docs, there are some limitations of the vision model that we should be aware of:

Medical images: The model is not suitable for interpreting specialized medical images like CT scans and shouldn't be used for medical advice.

Non-English: The model may not perform optimally when handling images with text of non-Latin alphabets, such as Japanese or Korean.

Small text: Enlarge text within the image to improve readability, but avoid cropping important details.

Rotation: The model may misinterpret rotated / upside-down text or images.

Visual elements: The model may struggle to understand graphs or text where colors or styles like solid, dashed, or dotted lines vary.

Spatial reasoning: The model struggles with tasks requiring precise spatial localization, such as identifying chess positions.

Accuracy: The model may generate incorrect descriptions or captions in certain scenarios.

Image shape: The model struggles with panoramic and fisheye images.

Metadata and resizing: The model doesn't process original file names or metadata, and images are resized before analysis, affecting their original dimensions.

Counting: May give approximate counts for objects in images.

CAPTCHAS: For safety reasons, we have implemented a system to block the submission of CAPTCHAs.

Overall, I do think the ability to combine text and image input as part of of the same chat is a game changer! This could unlock a lot of scenarios which were not possible just with a single mode of input. Very excited to see what is next!

Hope you found the post useful!