LLMFeed / app.py
codelion's picture
Update app.py
86ea854 verified
raw
history blame
7.59 kB
import gradio as gr
from google import genai
from google.genai import types
from PIL import Image
from io import BytesIO
import base64
import os
import json
# Initialize the Google Generative AI client with the API key from environment variables
try:
api_key = os.environ['GEMINI_API_KEY']
except KeyError:
raise ValueError("Please set the GEMINI_API_KEY environment variable.")
client = genai.Client(api_key=api_key)
def generate_item(tag):
"""
Generate a single feed item consisting of text from Gemini LLM and an image from Imagen.
Args:
tag (str): The tag to base the content on.
Returns:
dict: A dictionary with 'text' (str) and 'image_base64' (str).
"""
# Generate text using Gemini LLM with JSON output
prompt = f"""
Generate a short, engaging TikTok-style caption about {tag}.
Return the response as a JSON object with a single key 'caption' containing the caption text.
Example: {{"caption": "Craving this yummy treat! 😍 #foodie"}}
Do not include additional commentary or options.
"""
text_response = client.models.generate_content(
model='gemini-2.5-flash-preview-04-17',
contents=[prompt]
)
# Parse JSON response to extract the caption
try:
response_json = json.loads(text_response.text.strip())
text = response_json['caption']
except (json.JSONDecodeError, KeyError):
text = f"Wow, {tag} is amazing! 😍 #{tag}" # Fallback caption
# Generate an image based on the tag, avoiding text
image_prompt = f"""
A vivid, high-quality visual scene representing {tag}, designed for a TikTok video.
The image should be colorful and engaging, with no text or letters included.
"""
image_response = client.models.generate_images(
model='imagen-3.0-generate-002',
prompt=image_prompt,
config=types.GenerateImagesConfig(
number_of_images=1,
aspect_ratio="9:16",
person_generation="DONT_ALLOW"
)
)
# Check if images were generated
if image_response.generated_images and len(image_response.generated_images) > 0:
generated_image = image_response.generated_images[0]
image = Image.open(BytesIO(generated_image.image.image_bytes))
else:
# Fallback to a placeholder image
image = Image.new('RGB', (300, 533), color='gray') # 9:16 aspect ratio
# Convert the image to base64
buffered = BytesIO()
image.save(buffered, format="PNG")
img_str = base64.b64encode(buffered.getvalue()).decode()
return {'text': text, 'image_base64': img_str}
def start_feed(tag):
"""
Start a new feed with the given tag by generating one initial item.
Args:
tag (str): The tag to generate content for.
Returns:
tuple: (current_tag, feed_items, html_content)
"""
if not tag.strip():
tag = "trending" # Default tag if empty
item = generate_item(tag)
feed_items = [item]
html_content = generate_html(feed_items)
return tag, feed_items, html_content
def load_more(current_tag, feed_items):
"""
Append a new item to the existing feed using the current tag.
Args:
current_tag (str): The tag currently being used for the feed.
feed_items (list): The current list of feed items.
Returns:
tuple: (current_tag, updated_feed_items, updated_html_content)
"""
new_item = generate_item(current_tag)
feed_items.append(new_item)
html_content = generate_html(feed_items)
return current_tag, feed_items, html_content
def generate_html(feed_items):
"""
Generate an HTML string to display the feed items in a TikTok-like vertical layout.
Args:
feed_items (list): List of dictionaries containing 'text' and 'image_base64'.
Returns:
str: HTML string representing the feed.
"""
html_str = """
<div style="
display: flex;
flex-direction: column;
align-items: center;
max-width: 360px;
margin: 0 auto;
background-color: #000;
height: 640px;
overflow-y: auto;
scrollbar-width: none;
-ms-overflow-style: none;
border: 1px solid #333;
border-radius: 10px;
">
"""
# Hide scrollbar for a cleaner look
html_str += """
<style>
div::-webkit-scrollbar {
display: none;
}
</style>
"""
for item in feed_items:
html_str += f"""
<div style="
width: 100%;
height: 640px;
position: relative;
display: flex;
flex-direction: column;
justify-content: flex-end;
overflow: hidden;
">
<img src="data:image/png;base64,{item['image_base64']}" style="
width: 100%;
height: 100%;
object-fit: cover;
position: absolute;
top: 0;
left: 0;
z-index: 1;
">
<div style="
position: relative;
z-index: 2;
background: linear-gradient(to top, rgba(0,0,0,0.7), transparent);
padding: 20px;
color: white;
font-family: Arial, sans-serif;
font-size: 18px;
font-weight: bold;
text-shadow: 1px 1px 2px rgba(0,0,0,0.5);
">
{item['text']}
</div>
</div>
"""
html_str += "</div>"
return html_str
# Define the Gradio interface
with gr.Blocks(
css="""
body { background-color: #000; color: #fff; font-family: Arial, sans-serif; }
.gradio-container { max-width: 400px; margin: 0 auto; padding: 10px; }
input, select, button { border-radius: 5px; background-color: #222; color: #fff; border: 1px solid #444; }
button { background-color: #ff2d55; border: none; }
button:hover { background-color: #e0264b; }
.gr-button { width: 100%; margin-top: 10px; }
.gr-form { background-color: #111; padding: 15px; border-radius: 10px; }
""",
title="TikTok-Style Infinite Feed"
) as demo:
# Input section
with gr.Column(elem_classes="gr-form"):
gr.Markdown("### Create Your TikTok Feed")
with gr.Row():
suggested_tags = gr.Dropdown(
choices=["food", "travel", "fashion", "tech"],
label="Pick a Tag",
value="food"
)
tag_input = gr.Textbox(
label="Or Enter a Custom Tag",
value="food",
placeholder="e.g., sushi, adventure"
)
with gr.Row():
start_button = gr.Button("Start Feed")
load_more_button = gr.Button("Load More")
# Output display
feed_html = gr.HTML()
# State variables
current_tag = gr.State(value="")
feed_items = gr.State(value=[])
# Event handlers
def set_tag(selected_tag):
"""Update the tag input when a suggested tag is selected."""
return selected_tag
suggested_tags.change(fn=set_tag, inputs=suggested_tags, outputs=tag_input)
start_button.click(
fn=start_feed,
inputs=tag_input,
outputs=[current_tag, feed_items, feed_html]
)
load_more_button.click(
fn=load_more,
inputs=[current_tag, feed_items],
outputs=[current_tag, feed_items, feed_html]
)
# Launch the app
demo.launch()