Update app.py

app.py

@@ -1,28 +1,30 @@
-from transformers import load_tool, ReactCodeAgent, HfApiEngine
+from huggingface_hub import InferenceClient
+from langchain_huggingface import HuggingFaceHub
+from langchain.tools import DuckDuckGoSearchResults
+from langchain.agents import create_react_agent
+from langchain.tools import BaseTool
 from PIL import Image, ImageDraw, ImageFont
 import tempfile
 import gradio as gr
+import requests
+from io import BytesIO
+
+# Your HF API token here (set your actual token)
+#HF_TOKEN 
 
 #%% Methods
-# Function to add a label to an image
+
 def add_label_to_image(image, label):
-    # Create a drawing context
     draw = ImageDraw.Draw(image)
-    
-    # Define font size and color (adjust font path for your environment)
-    font_path = "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf"  # Example font path
-    font_size = 30  # Larger font size for better visibility
+    font_path = "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf"
+    font_size = 30
     try:
         font = ImageFont.truetype(font_path, font_size)
     except:
         font = ImageFont.load_default()
-    
-    # Calculate the size and position of the text (aligned to the left)
     text_bbox = draw.textbbox((0, 0), label, font=font)
     text_width, text_height = text_bbox[2] - text_bbox[0], text_bbox[3] - text_bbox[1]
-    position = (image.width - text_width - 20, image.height - text_height - 20)# right-aligned with margin
-    
-    # Add a semi-transparent rectangle behind the text for better visibility
+    position = (image.width - text_width - 20, image.height - text_height - 20)
     rect_margin = 10
     rect_position = [
         position[0] - rect_margin,
@@ -30,108 +32,129 @@ def add_label_to_image(image, label):
         position[0] + text_width + rect_margin,
         position[1] + text_height + rect_margin,
     ]
-    draw.rectangle(rect_position, fill=(0, 0, 0, 128))  # Semi-transparent black
+    draw.rectangle(rect_position, fill=(0, 0, 0, 128))
     draw.text(position, label, fill="white", font=font)
     return image
 
 
-# Function to plot, label, and save an image
 def plot_and_save_agent_image(agent_image, label, save_path=None):
-    # Convert AgentImage to a raw PIL Image
-    pil_image = agent_image.to_raw()
+    # agent_image is a PIL Image already in this refactor
+    pil_image = agent_image
 
-    # Add a label to the image
     labeled_image = add_label_to_image(pil_image, label)
-
-    # Plot the image using PIL's show method
     labeled_image.show()
 
-    # If save_path is provided, save the image
     if save_path:
         labeled_image.save(save_path)
         print(f"Image saved to {save_path}")
     else:
         print("No save path provided. Image not saved.")
 
-# Function to generate prompts for an object
+
 def generate_prompts_for_object(object_name):
-    prompts = {
+    return {
         "past": f"Show an old version of a {object_name} from its early days.",
         "present": f"Show a {object_name} with current features/design/technology.",
         "future": f"Show a futuristic version of a {object_name}, by predicting advanced features and futuristic design."
     }
-    return prompts
 
-# Function to generate the object's history images and GIF
+
 def generate_object_history(object_name):
     images = []
-    
-    # Get prompts for the object
     prompts = generate_prompts_for_object(object_name)
     labels = {
         "past": f"{object_name} - Past",
         "present": f"{object_name} - Present",
         "future": f"{object_name} - Future"
     }
-    
-    # Generate sequential images and display them
-    for time_period, frame in prompts.items():
-        print(f"Generating {time_period} frame: {frame}")
-        result = agent.run(frame)  # The tool generates the image
-        
-        # Append the image to the list for GIF creation
-        images.append(result.to_raw())  # Ensure we're using raw image for GIF
 
-        # Save each image with the appropriate name and label
-        image_filename = f"{object_name}_{time_period}.png"
-        plot_and_save_agent_image(result, labels[time_period], save_path=image_filename)
+    for time_period, prompt in prompts.items():
+        print(f"Generating {time_period} frame: {prompt}")
+        result = agent.invoke(prompt)  # returns PIL Image or string output
         
-    # Create GIF from images
+        # result is a PIL Image from our tool, or fallback string - ensure PIL Image
+        if isinstance(result, Image.Image):
+            images.append(result)
+            image_filename = f"{object_name}_{time_period}.png"
+            plot_and_save_agent_image(result, labels[time_period], save_path=image_filename)
+        else:
+            print(f"Unexpected output for {time_period}: {result}")
+
     gif_path = f"{object_name}_evolution.gif"
-    images[0].save(
-        gif_path, 
-        save_all=True, 
-        append_images=images[1:], 
-        duration=1000,  # Duration in milliseconds for each frame
-        loop=0          # Infinite loop
-    )
-    
-    # Return images and GIF path
+    if images:
+        images[0].save(
+            gif_path,
+            save_all=True,
+            append_images=images[1:],
+            duration=1000,
+            loop=0
+        )
+        print(f"GIF saved to {gif_path}")
+    else:
+        print("No images generated, GIF not created.")
+
     return images, gif_path
 
+
 #%% Initialization of tools and AI_Agent
-# Import text-to-image tool from Hub
-image_generation_tool = load_tool("m-ric/text-to-image", cache=False)
 
-# Import search tool from LangChain
-from transformers.agents.search import DuckDuckGoSearchTool
-search_tool = DuckDuckGoSearchTool()
+# Initialize HuggingFace Inference Client for text-to-image
+text_to_image_client = InferenceClient(repo_id="m-ric/text-to-image")
 
-# Load the LLM engine
-llm_engine = HfApiEngine("Qwen/Qwen2.5-72B-Instruct")
+def run_text_to_image(prompt: str) -> Image.Image:
+    outputs = text_to_image_client.text_to_image(prompt)
+    # Assuming outputs returns a list of URLs
+    image_url = outputs[0] if outputs else None
+    if image_url is None:
+        raise ValueError("No image URL returned from the model.")
+    response = requests.get(image_url)
+    img = Image.open(BytesIO(response.content)).convert("RGB")
+    return img
 
-# Initialize the agent with both tools
-agent = ReactCodeAgent(tools=[image_generation_tool, search_tool], llm_engine=llm_engine)
+# Custom LangChain tool wrapper for text-to-image
+class TextToImageTool(BaseTool):
+    name = "text-to-image"
+    description = "Generates an image from a prompt using HuggingFace model"
 
-# Gradio interface
+    def _run(self, prompt: str):
+        return run_text_to_image(prompt)
+
+    async def _arun(self, prompt: str):
+        raise NotImplementedError()
+
+image_generation_tool = TextToImageTool()
+
+# DuckDuckGo Search Tool from LangChain
+search_tool = DuckDuckGoSearchResults()
+
+# HuggingFace LLM for Qwen2.5
+llm_engine = HuggingFaceHub(
+    repo_id="Qwen/Qwen2.5-72B-Instruct",
+    huggingfacehub_api_token=HF_TOKEN,
+    model_kwargs={"temperature": 0.7}
+)
+
+# Create agent with the tools and LLM
+agent = create_react_agent(llm_engine, tools=[image_generation_tool, search_tool])
+
+
+#%% Gradio interface
 def create_gradio_interface():
     with gr.Blocks() as demo:
         gr.Markdown("# TimeMetamorphy: an object Evolution Generator")
         
-        # Add a section for instructions
         gr.Markdown("""
         ## Unlocking the secrets of time!
         This app unveils these mysteries by offering a unique/magic lens that allows us "time travel". 
         Powered by AI agents equipped with cutting-edge tools, it provides the superpower to explore the past, witness the present, and dream up the future like never before.
-        
+
         This system allows you to generate visualizations of how an object/concept, like a bicycle or a car, may have evolved over time. 
         It generates images of the object in the past, present, and future based on your input.
-        
+
         ### Default Example: Evolution of a Car
         Below, you can see a precomputed example of a "car" evolution. Enter another object to generate its evolution.
         """)
 
-        # Paths to the precomputed files
         default_images = [
             ("car_past.png", "Car - Past"),
             ("car_present.png", "Car - Present"),
@@ -141,25 +164,22 @@ def create_gradio_interface():
 
         with gr.Row():
             with gr.Column():
-                # Textbox for user to input an object name
-                object_name_input = gr.Textbox(label="Enter an object name (e.g., bicycle, phone)", 
-                                              placeholder="Enter an object name", 
-                                              lines=1)
-
-                # Button to trigger the generation of images and GIF
+                object_name_input = gr.Textbox(
+                    label="Enter an object name (e.g., bicycle, phone)", 
+                    placeholder="Enter an object name", 
+                    lines=1
+                )
                 generate_button = gr.Button("Generate Evolution")
-    
-                # Gradio Gallery component to display the images
-                image_gallery = gr.Gallery(label="Generated Images", show_label=True, columns=3, rows=1, value=default_images)
-
-                # Output for the generated GIF
+                image_gallery = gr.Gallery(
+                    label="Generated Images", show_label=True, columns=3, rows=1, value=default_images
+                )
                 gif_output = gr.Image(label="Generated GIF", show_label=True, value=default_gif_path)
-         
-        # Set the action when the button is clicked
+
         generate_button.click(fn=generate_object_history, inputs=[object_name_input], outputs=[image_gallery, gif_output])
     
     return demo
 
-# Launch the Gradio app 
+
+# Launch the Gradio app
 demo = create_gradio_interface()
-demo.launch(share=True)
+demo.launch(share=True)