How to load a local image to gpt4 -vision using API

I am not sure how to load a local image file to the gpt-4 vision.
Can someone explain how to do it?

from openai import OpenAI

client = OpenAI()

import matplotlib.image as mpimg
img123 = mpimg.imread('img.png')

response =
      "role": "user",
      "content": [
        {"type": "text", "text": "What’s in this image?"},
          "type": "image_url",
          "image_url": {
            "url" : "img123", 


I see that you are attempting to write a python script. I occasionally enjoy writing Python also.

I hope that meets your needs and informs where next to go with your own endeavors.

import os

This code is for v1 of the openai package: openai · PyPI

pip install openai

pip install requests

from openai import OpenAI
import base64
import requests

my_api_key = os.environ[“OPENAI_API_KEY”]

Function to encode the image

def encode_image(image_path):
with open(image_path, “rb”) as image_file:
return base64.b64encode(‘utf-8’)

Path to your image

image_path = “.\SourceImages\some_text.jpg”

Getting the base64 string

base64_image = encode_image(image_path)

Loads a local image file and OCRs it.

headers = {
“Content-Type”: “application/json”,
“Authorization”: f"Bearer {my_api_key}"

payload = {
“model”: “gpt-4-vision-preview”,
“messages”: [
“role”: “user”,
“content”: [
“type”: “text”,
“text”: “What’s in this image?”
“type”: “image_url”,
“image_url”: {
“url”: f"data:image/jpeg;base64,{base64_image}"
“max_tokens”: 300


response ="", headers=headers, json=payload)

except Exception as ex:
print(“Exception:”, ex)

How to perform a gpt-4-vision-preview prompt using the openai python module?

Here, the instructions for installing openai module were provided, but the OpenAI API call was made using curl ( How to use an openai functon call to do so? Can not find any documentation on this. Thanks

In response to this post, I spent a good amount of time coming up with the uber-example of using the gpt-4-vision model to send local files.

Stuff that doesn’t work in vision, so stripped:

  • functions
  • tools
  • logprobs
  • logit_bias


  • Local files: you store and send instead of relying on OpenAI fetch;
  • creating user message with base64 from files, upsampling and resizing, for multiple images per message;
  • calling with streaming in “raw” mode;
  • parsing, loading headers and displaying chunks as received.

This linear script serves as a basic example. Going beyond the scope of “example” would be considerations of methods like adaptive resizing based on content type, incorporating an additional parameter to accept memory images or URLs alongside files, and including a chat history with variable image expiration, among other things.

Documented by code:
undocumented base64-only endpoint API method

## supporting functions
import base64, textwrap, time, openai, os, io
from PIL import Image  # Pillow image library

def resize_image(image, max_dimension):
    width, height = image.size

    # Check if the image has a palette and convert it to true color mode
    if image.mode == "P":
        if "transparency" in
            image = image.convert("RGBA")
            image = image.convert("RGB")

    if width > max_dimension or height > max_dimension:
        if width > height:
            new_width = max_dimension
            new_height = int(height * (max_dimension / width))
            new_height = max_dimension
            new_width = int(width * (max_dimension / height))
        image = image.resize((new_width, new_height), Image.LANCZOS)
        timestamp = time.time()

    return image

def convert_to_png(image):
    with io.BytesIO() as output:, format="PNG")
        return output.getvalue()

def process_image(path, max_size):
    with as image:
        width, height = image.size
        mimetype = image.get_format_mimetype()
        if mimetype == "image/png" and width <= max_size and height <= max_size:
            with open(path, "rb") as f:
                encoded_image = base64.b64encode('utf-8')
                return (encoded_image, max(width, height))  # returns a tuple consistently
            resized_image = resize_image(image, max_size)
            png_image = convert_to_png(resized_image)
            return (base64.b64encode(png_image).decode('utf-8'),
                    max(width, height)  # same tuple metadata

def create_image_content(image, maxdim, detail_threshold):
    detail = "low" if maxdim < detail_threshold else "high"
    return {
        "type": "image_url",
        "image_url": {"url": f"data:image/jpeg;base64,{image}", "detail": detail}

def set_system_message(sysmsg):
    return [{
        "role": "system",
        "content": sysmsg

## user message with images function
def set_user_message(user_msg_str,
                     file_path_list=[],      # A list of file paths to images.
                     max_size_px=1024,       # Shrink images for lower expense
                     file_names_list=None,   # You can set original upload names to show AI
                     tiled=False,            # True is the API Reference method
                     detail_threshold=700):  # any images below this get 512px "low" mode

    if not isinstance(file_path_list, list):  # create empty list for weird input
        file_path_list = []

    if not file_path_list:  # no files, no tiles
        tiled = False

    if file_names_list and len(file_names_list) == len(file_path_list):
        file_names = file_names_list
        file_names = [os.path.basename(path) for path in file_path_list]

    base64_images = [process_image(path, max_size_px) for path in file_path_list]

    uploaded_images_text = ""
    if file_names:
        uploaded_images_text = "\n\n---\n\nUploaded images:\n" + '\n'.join(file_names)

    if tiled:
        content = [{"type": "text", "text": user_msg_str + uploaded_images_text}]
        content += [create_image_content(image, maxdim, detail_threshold)
                    for image, maxdim in base64_images]
        return [{"role": "user", "content": content}]
        return [{
            "role": "user",
            "content": ([user_msg_str + uploaded_images_text]
                        + [{"image": image} for image, _ in base64_images])
# -- START -- set up run variables

system_msg = """
You are VisionPal, an AI assistant powered by GPT-4 with computer vision.
AI knowledge cutoff: April 2023

Built-in vision capabilities:
- extract text from image
- describe images
- analyze image contents
- logical problem-solving requiring machine vision

# The user message
user_msg = """
How many images were received?
Describe the contents.
Describe the quality.
Repeat back the file names sent.

# user images file list, and max dimension limit
max_size = 1024  # downsizes if any dimension above this
image_paths = ["./img1.png", "./img2.png"]  # empty for no images
true_files = ["real file name 1.png", "real file name 2.jpg"]
true_files = None  # you can give real names if using temp upload locations
# Assemble the request parameters (all are dictionaries)
system = set_system_message(system_msg)
chat_hist = []  # list of more user/assistant items
user = set_user_message(user_msg, image_paths, max_size, file_names_list=true_files)

params = {  # dictionary format for ** unpacking
  "model": "gpt-4-vision-preview", "temperature": 0.5, "user": "my_customer",
  "max_tokens": 500, "top_p": 0.5, "stream": True,
  "messages": system + chat_hist + user,

start = time.perf_counter()
    client = openai.Client(timeout=111)
    response =**params)
    headers_dict = response.headers.items().mapping.copy()
    for key, value in headers_dict.items():  # set a variable for each header
        locals()[f'headers_{key.replace("-", "_")}'] = value
except Exception as e:
    print(f"Error during API call: {e}")
    response = None

if response is not None:
        reply = ""
        for chunk_no, chunk in enumerate(response.parse()):
            if chunk.choices[0].delta.content:
                reply += chunk.choices[0].delta.content
                print(chunk.choices[0].delta.content, end="")
    except Exception as e:
        print(f"Error during receive/parsing: {e}")

print(f"\n[elapsed: {time.perf_counter()-start:.2f} seconds]")

(Writing comments, docstrings, hinting, typing, education, etc. is where it becomes consulting work…)