Image upload for GPT-4V? (Feature request)

[mkammes]

GPT Plus members can use the upload function for media files, such as images.

https://platform.openai.com/docs/guides/vision/uploading-base-64-encoded-images

As an example, this would be great to have an automation to take a picture of your bar (or refrigerator) and provide recipe suggestions based on the ingredients....along with a custom prompt.

Here is an example: https://m.facebook.com/groups/HomeAssistant/permalink/3611503665787644/?mibextid=Nif5oz .

I've tried doing this via Python and pyscript with little success.

Thanks!

[jekalmin]

Thanks for a suggestion.

I just read the post on facebook, and it is really interesting feature. There are several things I would have to check before I'm certain that it is possible.

Details

1. The content type should be changed

currently messages form like below.

[
    {'role': 'system', 'content': "..."},
    {'role': 'user', 'content': 'turn on bedroom light'},
    {'role': 'function', 'name': 'execute_services', 'content': '[True]'}
]

In the openai guide you referenced, the content of message should be changed from string to list.

[
    {
        'role': 'user',
        'content': [
            {'type': 'text', 'text': 'What is in this image?'},
            {'type': 'image_url', 'image_url': 'https://...'}
        ]
    }
]

2. How to attach image_url

Since it's hard for the component(extended_openai_conversation) to attach "image_url" in user role, the only way, I can think of, is to provide a function that attaches "image_url" via function response.

I hope the format like below works, but there is no example that uses both function and image.

[
    {
        'role': 'function',
        'content': [
            {/* function response (don't know how this object will be formatted) */},
            {'type': 'image_url', 'image_url': 'https://...'}
        ]
    }
]

I will look into this when I have time and GPT Plus is resumed.

[mkammes]

Outstanding! I can confirm it works (as I've replicated it via my own python script via the OpenAI example).

I'm happy to create an API key for you to test with.

Thanks!

[jekalmin]

Thank you! I will try by myself first, and then ask for help if needed!

[jekalmin]

I just upgraded and tried gpt-4-vision-preview model. Unfortunately, it seems that this model doesn't support functions. I got an error like below

Code

functions = [
    {
        "name": "get_image",
        "description": "Get image",
        "parameters": {
            "type": "object",
            "properties": {
                "url": {
                    "type": "string",
                    "description": "The image url",
                }
            },
            "required": ["url"],
        },
    }
]


response = openai.ChatCompletion.create(
  model="gpt-4-vision-preview",
  functions=functions,
  function_call="auto",
  messages=[
    {
      "role": "user",
      "content": "What’s in an image?"
    }
  ],
  max_tokens=300,
)

print(response)

Request

DEBUG:openai:api_version=None data='{"model": "gpt-4-vision-preview", "functions": [{"name": "get_image", "description": "Get image", "parameters": {"type": "object", "properties": {"url": {"type": "string", "description": "The image url"}}, "required": ["url"]}}], "function_call": "auto", "messages": [{"role": "user", "content": "What\\u2019s in an image?"}], "max_tokens": 300}' message='Post details'

Response

openai.error.InvalidRequestError: 2 validation errors for Request
body -> function_call
  extra fields not permitted (type=value_error.extra)
body -> functions
  extra fields not permitted (type=value_error.extra)

Maybe I will add a service, so that you can hook it via functions

[jekalmin]

I have added "query_image" service in https://github.com/jekalmin/extended_openai_conversation/pull/60.

You can try adding function like below

Function

- spec:
    name: get_refrigerator_items
    description: Get description of items in refrigerator
    parameters:
      type: object
      properties:
        url:
          type: string
          description: image url of refrigerator
          enum:
            - https://i.pinimg.com/originals/8b/cc/f1/8bccf14daf77ce887fc162934335cb21.jpg # needs to change
      required:
      - url
  function:
    type: composite
    sequence:
      - type: script
        sequence:
          - service: extended_openai_conversation.query_image
            data:
              prompt: What alcohol and brands do you see in this picture?
              images:
                - url: "{{url}}"
              max_tokens: 300
              config_entry: YOUR_CONFIG_ENTRY_KEY # needs to change
            response_variable: _function_result
        response_variable: image_result
      - type: template
        value_template: "{{image_result.choices[0].message.content}}"

Then ask "what's in refrigerator"

[mkammes]

Outstanding! Great work. I look forward to testing this out!

[jekalmin]

Released this in 1.0.1-beta2