app.py

import os
import json
import numpy as np
import torch
from PIL import Image, ImageDraw
import gradio as gr
from openai import OpenAI
from geopy.geocoders import Nominatim
from staticmap import StaticMap, CircleMarker, Polygon
from diffusers import ControlNetModel, StableDiffusionControlNetInpaintPipeline
import spaces

# Initialize APIs
openai_client = OpenAI(api_key=os.environ['OPENAI_API_KEY'])
geolocator = Nominatim(user_agent="geoapi")

# Function to fetch coordinates
@spaces.GPU
def get_geo_coordinates(location_name):
    try:
        location = geolocator.geocode(location_name)
        if location:
            return [location.longitude, location.latitude]
        return None
    except Exception as e:
        print(f"Error fetching coordinates for {location_name}: {e}")
        return None

# Function to process OpenAI chat response
@spaces.GPU
def process_openai_response(query):
    response = openai_client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
    {
      "role": "system",
      "content": [
        {
          "type": "text",
          "text": "\"input\": \"\"\"You are a skilled assistant answering geographical and historical questions. For each question, generate a structured output in JSON format, based on city names without coordinates. The response should include:\
Answer: A concise response to the question.\
Feature Representation: A feature type based on city names (Point, LineString, Polygon, MultiPoint, MultiLineString, MultiPolygon, GeometryCollection).\
Description: A prompt for a diffusion model describing the what should we draw regarding that.\
\
Handle the following cases:\
\
1. **Single or Multiple Points**: Create a point or a list of points for multiple cities.\
2. **LineString**: Create a line between two cities.\
3. **Polygon**: Represent an area formed by three or more cities (closed). Example: Cities forming a triangle (A, B, C).\
4. **MultiPoint, MultiLineString, MultiPolygon, GeometryCollection**: Use as needed based on the question.\
\
For example, if asked about cities forming a polygon, create a feature like this:\
\
Input: Mark an area with three cities.\
Output: {\"input\": \"Mark an area with three cities.\", \"output\": {\"answer\": \"The cities A, B, and C form a triangle.\", \"feature_representation\": {\"type\": \"Polygon\", \"cities\": [\"A\", \"B\", \"C\"], \"properties\": {\"description\": \"satelite image of a plantation, green fill, 4k, map, detailed, greenary, plants, vegitation, high contrast\"}}}}\
\
Ensure all responses are descriptive and relevant to city names only, without coordinates.\
\"}\"}"
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "draw a map in coconut triangle of sri lanka: The Coconut Triangle is a region in Sri Lanka that's known for its coconut production. It's made up of the districts of Kurunegala, Puttalam, and Gampaha."
        }
      ]
    }
  ],
        temperature=1,
        max_tokens=2048,
        top_p=1,
        frequency_penalty=0,
        presence_penalty=0,
        response_format={"type": "json_object"}
    )
    return json.loads(response.choices[0].message.content)

# Generate GeoJSON from OpenAI response
@spaces.GPU
def generate_geojson(response):
    feature_type = response['output']['feature_representation']['type']
    city_names = response['output']['feature_representation']['cities']
    properties = response['output']['feature_representation']['properties']

    coordinates = []
    for city in city_names:
        coord = get_geo_coordinates(city)
        if coord:
            coordinates.append(coord)

    if feature_type == "Polygon":
        coordinates.append(coordinates[0])  # Close the polygon

    return {
        "type": "FeatureCollection",
        "features": [{
            "type": "Feature",
            "properties": properties,
            "geometry": {
                "type": feature_type,
                "coordinates": [coordinates] if feature_type == "Polygon" else coordinates
            }
        }]
    }

# Generate static map image
@spaces.GPU
def generate_static_map(geojson_data):
    # Create a static map object with specified dimensions
    m = StaticMap(500, 500)
    
    # Process each feature in the GeoJSON
    for feature in geojson_data["features"]:
        geom_type = feature["geometry"]["type"]
        coords = feature["geometry"]["coordinates"]

        if geom_type == "Point":
            # Add a blue marker for Point geometries
            m.add_marker(CircleMarker((coords[0], coords[1]), 'blue', 10))
        elif geom_type in ["MultiPoint", "LineString"]:
            # Add a red marker for each point in MultiPoint or LineString geometries
            for coord in coords:
                m.add_marker(CircleMarker((coord[0], coord[1]), 'red', 10))
        elif geom_type in ["Polygon", "MultiPolygon"]:
            # Add green polygons for Polygon or MultiPolygon geometries
            for polygon in coords:
                m.add_polygon(Polygon([(c[0], c[1]) for c in polygon], 'green', 3))
    
    # Render the static map and return the Pillow Image object
    return m.render(zoom=10)

# ControlNet pipeline setup
controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16)
pipeline = StableDiffusionControlNetInpaintPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-inpainting", controlnet=controlnet, torch_dtype=torch.float16
)
# ZeroGPU compatibility
pipeline.to('cuda')

@spaces.GPU
def make_inpaint_condition(init_image, mask_image):
    init_image = np.array(init_image.convert("RGB")).astype(np.float32) / 255.0
    mask_image = np.array(mask_image.convert("L")).astype(np.float32) / 255.0

    assert init_image.shape[0:1] == mask_image.shape[0:1], "image and image_mask must have the same image size"
    init_image[mask_image > 0.5] = -1.0  # set as masked pixel
    init_image = np.expand_dims(init_image, 0).transpose(0, 3, 1, 2)
    init_image = torch.from_numpy(init_image)
    return init_image

@spaces.GPU
def generate_satellite_image(init_image, mask_image, prompt):
    control_image = make_inpaint_condition(init_image, mask_image)
    result = pipeline(
        prompt=prompt, 
        image=init_image, 
        mask_image=mask_image, 
        control_image=control_image,
        strength=0.65,
        guidance_scale=85
        )
    return result.images[0]

# Gradio UI
@spaces.GPU
def handle_query(query):
    # Process OpenAI response
    response = process_openai_response(query)
    geojson_data = generate_geojson(response)

    # Generate map image
    map_image = generate_static_map(geojson_data)

    # Generate mask for ControlNet
    empty_map = Image.new("RGB", map_image.size, "white")
    difference = np.array(map_image) - np.array(empty_map)
    mask = np.any(difference != 0, axis=-1).astype(np.uint8) * 255

    # Convert mask to PIL Image
    mask_image = Image.fromarray(mask)

    # Generate satellite image
    satellite_image = generate_satellite_image(map_image, mask_image, response['output']['feature_representation']['properties']['description'])

    return map_image, satellite_image, response

# Gradio interface
with gr.Blocks() as demo:
    with gr.Row():
        query_input = gr.Textbox(label="Enter Query")
        submit_btn = gr.Button("Submit")
    with gr.Row():
        map_output = gr.Image(label="Map Visualization")
        satellite_output = gr.Image(label="Generated Satellite Image")
        image_prompt = gr.Textbox(label="Image Prompt Used")
    submit_btn.click(handle_query, inputs=[query_input], outputs=[map_output, satellite_output, image_prompt])

if __name__ == "__main__":
    demo.launch()