api/index.py

import os
import base64
import numpy as np
from PIL import Image
import io
import requests

import replicate
from flask import Flask, request
import gradio as gr

import openai
from openai import OpenAI


from dotenv import load_dotenv, find_dotenv

import json


# Locate the .env file
dotenv_path = find_dotenv()

load_dotenv(dotenv_path)

OPENAI_API_KEY = os.getenv('OPENAI_API_KEY')
REPLICATE_API_TOKEN = os.getenv('REPLICATE_API_TOKEN')



client = OpenAI()


def call_openai(pil_image):
    # Save the PIL image to a bytes buffer
    buffered = io.BytesIO()
    pil_image.save(buffered, format="JPEG")
    
    # Encode the image to base64
    image_data = base64.b64encode(buffered.getvalue()).decode('utf-8')
    
    try:
        response = client.chat.completions.create(
            model="gpt-4o",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "You are a product designer. I've attached a moodboard here. In one sentence, what do all of these elements have in common? Answer from a design language perspective, if you were telling another designer to create something similar, including any repeating colors and materials and shapes and textures. This is for a single product, so respond as though you're applying them to a single object. Reply with a completion to the following (don't include these words please, just the rest): [A render of an object which] [your response]. Do NOT include 'A render of an object which' in your response."},
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": "data:image/jpeg;base64," + image_data,
                            },
                        },
                    ],
                }
            ],
            max_tokens=300,
        )
        return response.choices[0].message.content
    except openai.BadRequestError as e:
        print(e)
        print("e type")
        print(type(e))
        raise gr.Error(f"Please retry with a different moodboard file (below 20 MB in size and is of one the following formats: ['png', 'jpeg', 'gif', 'webp'])")
    except Exception as e:
        raise gr.Error("Unknown Error")


# Todo -- better prompt generator, add another LLM layer combining the user prompt and moodboard description (in the case of the jacket, 'high quality render of yellow jacket, its fabric is a pattern of cosmic etc etc' worked well)
# Could even do this 4 different times to get more diversity of renders
# Add "simple" to prompt before word

def image_classifier(moodboard, prompt):

    if moodboard is not None:
        pil_image = Image.fromarray(moodboard.astype('uint8'))

        openai_response = call_openai(pil_image)

    else:
        raise gr.Error(f"Please upload a moodboard to control image generation style")

    input = {
        "prompt": "high quality render of a " + prompt + " which " + openai_response + ", minimalist and simple mockup on a white background",
        "output_format": "jpg"
    }
    
    try:
        output = replicate.run(
            "stability-ai/stable-diffusion-3",
            input=input
        )
    except Exception as e:
        raise gr.Error(f"Error: {e}")

    try:
        image_url = output[0]
        response = requests.get(image_url)
        img1 = Image.open(io.BytesIO(response.content))
    except Exception as e:
        raise gr.Error(f"Image download failed: {e}")

    input["aspect_ratio"] = "3:2"
    input["cfg"] = 6
    
    try:
        output = replicate.run(
            "stability-ai/stable-diffusion-3",
            input=input
        )
        image_url = output[0]
        response = requests.get(image_url)
        img2 = Image.open(io.BytesIO(response.content))
    except Exception as e:
        raise gr.Error(f"Second image download failed: {e}")

    # Call SDXL API with the response from OpenAI
    input = {
        "width": 768,
        "height": 768,
        "prompt": "centered high quality render of a " + prompt + " which " + openai_response + ' centered on a plain white background',
        "negative_prompt": "worst quality, low quality, illustration, 2d, painting, cartoons, sketch, logo, buttons, markings, text, wires, complex, screws, nails, construction",
        "refine": "expert_ensemble_refiner",
        "apply_watermark": False,
        "num_inference_steps": 25,
        "num_outputs": 2,
        "guidance_scale": 8.5
    }
    
    output = replicate.run(
        "stability-ai/sdxl:7762fd07cf82c948538e41f63f77d685e02b063e37e496e96eefd46c929f9bdc",
        input=input
    )

    images = [img1, img2]

    for i in range(min(len(output), 2)):
        image_url = output[i]
        response = requests.get(image_url)
        images.append(Image.open(io.BytesIO(response.content)))
    
    # Add empty images if fewer than 3 were returned
    while len(images) < 4:
        images.append(Image.new('RGB', (768, 768), 'gray'))

    images.reverse()
    return images

demo = gr.Interface(fn=image_classifier, inputs=["image", "text"], outputs=["image", "image", "image", "image"])
demo.launch(share=True)