mostlycached's picture
Upload 3 files
0773ab4 verified
raw
history blame
13.2 kB
# app.py - TextDiffuser-2 implementation for Hugging Face Spaces
import os
import torch
import gradio as gr
import numpy as np
import json
from PIL import Image, ImageDraw, ImageFont
from transformers import AutoTokenizer, AutoModelForCausalLM
from diffusers import StableDiffusionPipeline
# Check for GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
class SimpleTextDiffuser:
"""
Simple implementation of TextDiffuser-2 concept for Hugging Face Spaces
"""
def __init__(self):
# Load language model for layout generation
# Using a small model for efficiency
self.tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
self.language_model = AutoModelForCausalLM.from_pretrained("distilgpt2")
self.language_model.to(device)
# Only load the diffusion model if we have a GPU
self.diffusion_model = None
if torch.cuda.is_available():
self.diffusion_model = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16
)
self.diffusion_model.to(device)
print("Models initialized")
def generate_layout(self, prompt, image_size=(512, 512), num_text_elements=3):
"""Generate text layout based on prompt"""
width, height = image_size
# Format the prompt for layout generation
layout_prompt = f"""
Create a layout for an image with:
- Description: {prompt}
- Image size: {width}x{height}
- Number of text elements: {num_text_elements}
Generate text content and positions:
"""
# Generate layout using LM
input_ids = self.tokenizer.encode(layout_prompt, return_tensors="pt").to(device)
with torch.no_grad():
output = self.language_model.generate(
input_ids,
max_length=input_ids.shape[1] + 150,
temperature=0.7,
num_return_sequences=1,
pad_token_id=self.tokenizer.eos_token_id
)
layout_text = self.tokenizer.decode(output[0], skip_special_tokens=True)
# Parse the generated layout (simplified)
# In a real implementation, this would be more sophisticated
text_elements = []
# Simple fallback: generate random layout
import random
# Create a title element
title = prompt.split()[:5]
title = " ".join(title) + "..."
title_x = width // 4
title_y = height // 4
text_elements.append({
"text": title,
"position": (title_x, title_y),
"size": 24,
"color": (0, 0, 0),
"type": "title"
})
# Create additional text elements
sample_texts = [
"Premium Quality",
"Best Value",
"Limited Edition",
"New Collection",
"Special Offer",
"Coming Soon",
"Best Seller",
"Top Choice",
"Featured Product",
"Exclusive Deal"
]
for i in range(1, num_text_elements):
x = random.randint(width // 8, width * 3 // 4)
y = random.randint(height // 3, height * 3 // 4)
text = sample_texts[i % len(sample_texts)]
color = (
random.randint(0, 200),
random.randint(0, 200),
random.randint(0, 200)
)
text_elements.append({
"text": text,
"position": (x, y),
"size": 18,
"color": color,
"type": f"element_{i}"
})
return text_elements, layout_text
def generate_image(self, prompt, image_size=(512, 512)):
"""Generate base image using diffusion model or placeholder"""
width, height = image_size
if self.diffusion_model and torch.cuda.is_available():
# Generate image using diffusion model
image = self.diffusion_model(
prompt=prompt,
height=height,
width=width,
num_inference_steps=30
).images[0]
else:
# Create a placeholder gradient image
image = Image.new("RGB", image_size, (240, 240, 240))
# Add a colored gradient background
for y in range(height):
for x in range(width):
r = int(240 - 100 * (y / height))
g = int(240 - 50 * (x / width))
b = int(240 - 75 * ((x + y) / (width + height)))
image.putpixel((x, y), (r, g, b))
return image
def render_text(self, image, text_elements):
"""Render text elements onto the image"""
image_with_text = image.copy()
draw = ImageDraw.Draw(image_with_text)
for element in text_elements:
try:
font_size = element["size"]
# Try to load a font, fall back to default if not available
try:
font = ImageFont.truetype("DejaVuSans.ttf", font_size)
except IOError:
try:
font = ImageFont.truetype("Arial.ttf", font_size)
except IOError:
font = ImageFont.load_default()
# Draw text with background for better visibility
text = element["text"]
position = element["position"]
color = element["color"]
# Get text size to create background
bbox = draw.textbbox(position, text, font=font)
text_width = bbox[2] - bbox[0]
text_height = bbox[3] - bbox[1]
# Draw semi-transparent background
padding = 5
background_box = [
position[0] - padding,
position[1] - padding,
position[0] + text_width + padding,
position[1] + text_height + padding
]
draw.rectangle(background_box, fill=(255, 255, 255, 200))
# Draw text
draw.text(position, text, fill=color, font=font)
except Exception as e:
print(f"Error rendering text: {e}")
continue
return image_with_text
def visualize_layout(self, text_elements, image_size=(512, 512)):
"""Create a visualization of the text layout"""
width, height = image_size
image = Image.new("RGB", image_size, (255, 255, 255))
draw = ImageDraw.Draw(image)
# Draw grid
for x in range(0, width, 50):
draw.line([(x, 0), (x, height)], fill=(230, 230, 230))
for y in range(0, height, 50):
draw.line([(0, y), (width, y)], fill=(230, 230, 230))
# Draw text elements
for element in text_elements:
position = element["position"]
text = element["text"]
element_type = element.get("type", "unknown")
# Draw position marker
circle_radius = 5
circle_bbox = [
position[0] - circle_radius,
position[1] - circle_radius,
position[0] + circle_radius,
position[1] + circle_radius
]
draw.ellipse(circle_bbox, fill=(255, 0, 0))
# Draw text label
try:
font = ImageFont.truetype("DejaVuSans.ttf", 12)
except IOError:
font = ImageFont.load_default()
# Draw text preview and position info
info_text = f"{text} ({element_type})"
pos_text = f"Position: ({position[0]}, {position[1]})"
draw.text((position[0] + 10, position[1]), info_text, fill=(0, 0, 0), font=font)
draw.text((position[0] + 10, position[1] + 15), pos_text, fill=(0, 0, 255), font=font)
return image
def generate_text_image(self, prompt, width=512, height=512, num_text_elements=3):
"""Generate an image with rendered text based on prompt"""
# Validate inputs
width = max(256, min(1024, width))
height = max(256, min(1024, height))
num_text_elements = max(1, min(5, num_text_elements))
image_size = (width, height)
# Step 1: Generate text layout
text_elements, layout_text = self.generate_layout(prompt, image_size, num_text_elements)
# Step 2: Generate base image
base_image = self.generate_image(prompt, image_size)
# Step 3: Render text onto the image
image_with_text = self.render_text(base_image, text_elements)
# Step 4: Create layout visualization
layout_visualization = self.visualize_layout(text_elements, image_size)
# Step 5: Format layout information for display
layout_info = {
"prompt": prompt,
"image_size": image_size,
"num_text_elements": num_text_elements,
"text_elements": text_elements,
"layout_generation_prompt": layout_text
}
formatted_layout = json.dumps(layout_info, indent=2)
return image_with_text, layout_visualization, formatted_layout
# Initialize the model
model = SimpleTextDiffuser()
# Define the Gradio interface
def process_request(prompt, width, height, num_text_elements):
try:
width = int(width)
height = int(height)
num_text_elements = int(num_text_elements)
image, layout, layout_info = model.generate_text_image(
prompt,
width=width,
height=height,
num_text_elements=num_text_elements
)
return image, layout, layout_info
except Exception as e:
error_message = f"Error: {str(e)}"
print(error_message)
return None, None, error_message
# Create the Gradio app
with gr.Blocks(title="TextDiffuser-2 Demo") as demo:
gr.Markdown("""
# TextDiffuser-2 Demo
This demo implements the concepts from the paper "[TextDiffuser-2: Unleashing the Power of Language Models for Text Rendering](https://arxiv.org/abs/2311.16465)" by Jingye Chen et al.
Generate images with text by providing a descriptive prompt below.
""")
with gr.Row():
with gr.Column(scale=1):
prompt_input = gr.Textbox(
label="Prompt",
value="A modern business poster with company name and tagline",
lines=3
)
with gr.Row():
width_input = gr.Number(label="Width", value=512, minimum=256, maximum=1024, step=64)
height_input = gr.Number(label="Height", value=512, minimum=256, maximum=1024, step=64)
num_elements_input = gr.Slider(
label="Number of Text Elements",
minimum=1,
maximum=5,
value=3,
step=1
)
submit_button = gr.Button("Generate Image", variant="primary")
with gr.Column(scale=2):
with gr.Tabs():
with gr.TabItem("Generated Image"):
image_output = gr.Image(label="Image with Text")
with gr.TabItem("Layout Visualization"):
layout_output = gr.Image(label="Text Layout")
with gr.TabItem("Layout Information"):
layout_info_output = gr.Code(language="json", label="Layout Data")
gr.Markdown("""
## Example Prompts
Try these prompts or create your own:
""")
examples = gr.Examples(
examples=[
["A movie poster for a sci-fi thriller", 512, 768, 3],
["A motivational quote on a sunset background", 768, 512, 2],
["A coffee shop menu with prices", 512, 512, 4],
["A modern business card design", 512, 384, 3],
],
inputs=[prompt_input, width_input, height_input, num_elements_input]
)
submit_button.click(
fn=process_request,
inputs=[prompt_input, width_input, height_input, num_elements_input],
outputs=[image_output, layout_output, layout_info_output]
)
gr.Markdown("""
## About
This is a simplified implementation for demonstration purposes. The full approach described in the paper involves deeper integration of language models with the diffusion process.
Running on: """ + str(device))
# Launch the app
if __name__ == "__main__":
demo.launch()