GPT-Vision & GPT

The GPT class is designed for text-to-text interactions, utilizing models from the Generative Pre-trained Transformer series. These models are adept at understanding and generating text based on the input provided.

​

Initialization

GPT(api_key="sk-...", model="gpt-4-turbo")

​

Methods:

• GPT.prompt(query: str) -> str

  Takes a text query and returns the model’s text response.

The GPTVision class extends the capabilities of GPT to include handling image inputs, making it suitable for image-to-text tasks. This class is part of the multimodal models that can process both text and images.

​

Initialization:

​

Methods:

• GPTVision.prompt(query: str) -> str

  Sends the prompt and returns the respone

• GPTVision.prompt_with_image(image: Image.Image, query: str) -> str

  Sends the prompt and image to the model and returns the response.

• GPTVision.parse_text(image: Image.Image) -> str

  Returns all text extracted from an image

​

When to use

GPT-4o and GPT-4v are the leading Vision Language models and are capable of understanding relatively fine details in images. Although when scaling to datasets of millions of images, these models can become prohibitively expensive.

Generally, GPT-4o is the best model to use for image-to-text tasks, and is cheaper than GPT-4v.

​

Usage Scenarios

wget -O drivers_license.jpg https://github.com/overeasy-sh/overeasy/blob/main/examples/drivers_license.jpg?raw=true

from overeasy import *
from pydantic import BaseModel
from PIL import Image

class License(BaseModel):
    license_number: str
    expiration_date: str
    state: str
    data_of_birth: str
    address: str
    first_name: str
    last_name: str
    
workflow = Workflow([
    InstructorImageAgent(response_model=License, model="gpt-4o")
])

image = Image.open("./drivers_license.jpg")
result, graph = workflow.execute(image)
print(repr(result[0].data))

License(
    license_number='I1234568', 
    expiration_date='08/31/2014', 
    state='California', 
    data_of_birth='08/31/1977', 
    address='2570 24th Street, Anytown, CA 95818', 
    first_name='Ima', 
    last_name='Cardholder'
)

wget -O eiffel_photo.png https://github.com/overeasy-sh/overeasy/blob/main/examples/eiffel_photo.png?raw=true

from overeasy import *
from PIL import Image

workflow = Workflow([
    DenseCaptioningAgent(model=GPTVision(model="gpt-4o"))
])
result, graph = workflow.execute(Image.open("./eiffel.png"))
print(result[0].data)

wget ...

from overeasy import *
from PIL import Image

workflow = Workflow([
    BinaryChoiceAgent("Do the Celtics have possession of the ball?", model=GPTVision(model="gpt-4o"))
])
image=Image.open("bball.png")
result, graph = workflow.execute(image)
print(result[0].data.class_names)

​

Limitations

Many GPT Visual models have relatively strict guard rails around inputs/photos of people so your requests may get screened out if they are related to these topics. These models also generally struggle with spatial reasoning(i.e. is object A in front of object B?).

For more details check the OpenAI docs