GPT-4-Vision-Preview fidelity/detail-parameter

API
1

The API documentation mentions a detail parameter (low or high) which controls the resolution in which the model views the image. But its not clear how to pass that parameter in a request. Im using python and the openai package. This is a code snippet from my ai-class. Where would I add the detail parameter?

    def askAboutImage(self,npImg,prompt):
        messages = [{"role":"user","content":[{"type":"text","text":prompt},
                                             {"type":"image_url","image_url":f"data:image/jpeg;base64,{np2base64(npImg)}"}]}]
        response = self.client.chat.completions.create(model=self.model,
                                                       messages = messages,
                                                       max_tokens=300)
        print(response.choices[0])
4 Likes
2

This is totally unclear on the documentation, I’m looking everywhere and can’t find the answer. I don’t see any argument called “detail” anywhere in the code

1 Like
3

You can add the parameter like:

image_url: {
    url: "https://your-image-url.com",
    detail: "low",
}
4 Likes
4

@WildRoot It works! Thank you!!

5

What’s the default if not specified low or high?

6

My guess would be low, if I get chance, I’ll take a look at the API library code on git.

7

I just ran the same prompt with the same image input twice.The first time I specified ‘low’, which cost me 230 tokens and without specification it cost 1700. So its save to assume that the default is ‘high’.

8

Have you tried to compare without specification to high? I think that’s the last test to confirm.

9

Yes, by now I ran all 3 options:
‘high’
‘low’
without specification

I can confirm that the default is high.

1 Like
10

The default is auto, which means if the image is of low resolution then it converts it into 512*512 and processes it using 65 tokens, else it sets to high and uses 130 tokens