hidream-arena / app.py
chengzeyi's picture
fix
b3b6622
raw
history blame
29.5 kB
import gradio as gr
import asyncio
import aiohttp
import time
from datetime import datetime
import plotly.graph_objects as go
from typing import Dict, List
import os
from dotenv import load_dotenv
import json
from PIL import Image, ImageDraw, ImageFont
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import uuid
import threading
import functools
# Load environment variables first
load_dotenv()
# Constants
API_BASE_URL = "https://api.wavespeed.ai/api/v2"
API_KEY = os.getenv("WAVESPEED_API_KEY") # Move API_KEY to global scope
if not API_KEY:
raise ValueError("WAVESPEED_API_KEY not found in environment variables")
# Rest of constants
BACKENDS = {
"flux-dev": {
"endpoint": f"{API_BASE_URL}/wavespeed-ai/flux-dev",
"name": "Flux-dev",
"color": "#FF9800",
},
"hidream-dev": {
"endpoint": f"{API_BASE_URL}/wavespeed-ai/hidream-i1-dev",
"name": "HiDream-dev",
"color": "#2196F3",
},
"hidream-full": {
"endpoint": f"{API_BASE_URL}/wavespeed-ai/hidream-i1-full",
"name": "HiDream-full",
"color": "#4CAF50",
},
}
MODEL_URL = "TostAI/nsfw-text-detection-large"
TITLE = "🖼️🔍 Image Prompt Safety Classifier 🛡️"
DESCRIPTION = "✨ Enter an image generation prompt to classify its safety level! ✨"
# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_URL)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_URL)
# Define class names with emojis and detailed descriptions
CLASS_NAMES = {
0: "✅ SAFE - This prompt is appropriate and harmless.",
1: "⚠️ QUESTIONABLE - This prompt may require further review.",
2: "🚫 UNSAFE - This prompt is likely to generate inappropriate content."
}
@functools.lru_cache(maxsize=128)
def classify_text(text):
inputs = tokenizer(text,
return_tensors="pt",
truncation=True,
padding=True,
max_length=1024)
with torch.no_grad():
outputs = model(**inputs)
logits = outputs.logits
predicted_class = torch.argmax(logits, dim=1).item()
return predicted_class, CLASS_NAMES[predicted_class]
class BackendStatus:
def __init__(self):
self.reset()
self.history: List[Dict] = []
def reset(self):
self.status = "idle"
self.progress = 0
self.start_time = None
self.end_time = None
def start(self):
self.status = "processing"
self.progress = 0
self.start_time = time.time()
self.end_time = None
def complete(self):
self.status = "completed"
self.progress = 100
self.end_time = time.time()
self.history.append({
"timestamp": datetime.now(),
"duration": self.end_time - self.start_time
})
def fail(self):
self.status = "failed"
self.end_time = time.time()
class SessionManager:
_instances = {}
_lock = threading.Lock()
@classmethod
def get_manager(cls, session_id=None):
if session_id is None:
session_id = str(uuid.uuid4())
with cls._lock:
if session_id not in cls._instances:
cls._instances[session_id] = GenerationManager()
return session_id, cls._instances[session_id]
@classmethod
def cleanup_old_sessions(cls, max_age=3600): # 1 hour default
current_time = time.time()
with cls._lock:
to_remove = []
for session_id, manager in cls._instances.items():
if (hasattr(manager, "last_activity")
and current_time - manager.last_activity > max_age):
to_remove.append(session_id)
for session_id in to_remove:
del cls._instances[session_id]
class GenerationManager:
def __init__(self):
self.backend_statuses = {
backend: BackendStatus()
for backend in BACKENDS
}
self.last_activity = time.time()
self.request_timestamps = [] # Track timestamps of requests
def update_activity(self):
self.last_activity = time.time()
def add_request_timestamp(self):
self.request_timestamps.append(time.time())
def has_exceeded_limit(self,
limit=10): # Default limit: 10 requests per hour
current_time = time.time()
# Filter timestamps to only include those within the last hour
self.request_timestamps = [
ts for ts in self.request_timestamps if current_time - ts <= 3600
]
return len(self.request_timestamps) >= limit
def get_performance_plot(self):
fig = go.Figure()
has_data = False
for backend, status in self.backend_statuses.items():
durations = [h["duration"] for h in status.history]
if durations:
has_data = True
avg_duration = sum(durations) / len(durations)
# Use bar chart instead of box plot
fig.add_trace(
go.Bar(
y=[avg_duration], #
x=[BACKENDS[backend]["name"]], # Backend name
name=BACKENDS[backend]["name"],
marker_color=BACKENDS[backend]["color"],
text=[f"{avg_duration:.2f}s"], # Show time in seconds
textposition="auto",
width=[0.5], # Make bars narrower
))
# Set a minimum y-axis range if we have data
if has_data:
max_duration = max([
max([h["duration"] for h in status.history] or [0])
for status in self.backend_statuses.values()
])
# Add 20% padding to the top
y_max = max_duration * 1.2
# Ensure the y-axis always starts at 0
fig.update_yaxes(range=[0, y_max])
fig.update_layout(
title="Average Generation Time (Including Network Latency)",
yaxis_title="Seconds",
xaxis_title="",
showlegend=False,
template="simple_white",
height=400, # Increase height
margin=dict(l=50, r=50, t=50, b=50), # Add margins
font=dict(size=14), # Larger font
)
# Make sure we have a valid figure even if no data
if not has_data:
fig.add_annotation(
text="No timing data available yet",
xref="paper",
yref="paper",
x=0.5,
y=0.5,
showarrow=False,
font=dict(size=16),
)
return fig
async def submit_task(self, backend: str, prompt: str) -> str:
status = self.backend_statuses[backend]
status.start()
try:
url = BACKENDS[backend]["endpoint"]
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {API_KEY}",
}
payload = {
"prompt": prompt,
"enable_safety_checker": True,
"enable_base64_output": True, # Enable base64 output
"size": "1024*1024",
"seed": -1,
}
print(f"Submitting task to {backend}")
print(f"URL: {url}")
print(f"Payload: {json.dumps(payload, indent=2)}")
# Use aiohttp instead of requests for async
async with aiohttp.ClientSession() as session:
async with session.post(url, headers=headers,
json=payload) as response:
if response.status == 200:
result = await response.json()
request_id = result["data"]["id"]
print(
f"Task submitted successfully. Request ID: {request_id}"
)
return request_id
else:
text = await response.text()
raise Exception(
f"API error: {response.status}, {text}")
except Exception as e:
status.fail()
raise Exception(f"Failed to submit task: {str(e)}")
# Add this method to reset history
def reset_history(self):
"""Reset history for all backends"""
for status in self.backend_statuses.values():
status.history = [] # Clear history data
return self
class ClientManager:
_instances = {}
_lock = threading.Lock()
@classmethod
def get_manager(cls, client_id=None):
if not client_id:
client_id = str(uuid.uuid4())
with cls._lock:
if client_id not in cls._instances:
cls._instances[client_id] = ClientGenerationManager()
return cls._instances[client_id]
@classmethod
def cleanup_old_clients(cls, max_age=3600): # 1 hour default
current_time = time.time()
with cls._lock:
to_remove = []
for client_id, manager in cls._instances.items():
if (hasattr(manager, "last_activity")
and current_time - manager.last_activity > max_age):
to_remove.append(client_id)
for client_id in to_remove:
del cls._instances[client_id]
class ClientGenerationManager:
def __init__(self):
self.lock = threading.Lock()
self.last_activity = time.time()
self.request_timestamps = [] # Track timestamps of requests
def update_activity(self):
with self.lock:
self.last_activity = time.time()
def add_request_timestamp(self):
with self.lock:
self.request_timestamps.append(time.time())
def has_exceeded_limit(self, limit=20):
with self.lock:
current_time = time.time()
# Filter timestamps to only include those within the last hour
self.request_timestamps = [
ts for ts in self.request_timestamps
if current_time - ts <= 3600
]
return len(self.request_timestamps) >= limit
# Helper function to create error images as data URIs
def create_error_image(backend, error_message):
try:
import base64
from io import BytesIO
# Create an in-memory image
img = Image.new("RGB", (512, 512), color="#ffdddd")
draw = ImageDraw.Draw(img)
try:
font = ImageFont.truetype("Arial", 20)
except:
font = ImageFont.load_default()
# Wrap and draw error message
words = error_message.split(" ")
lines = []
line = ""
for word in words:
if len(line + word) < 40:
line += word + " "
else:
lines.append(line)
line = word + " "
if line:
lines.append(line)
y_position = 100
for line in lines:
draw.text((50, y_position), line, fill="black", font=font)
y_position += 30
# Save to a BytesIO object instead of a file
buffer = BytesIO()
img.save(buffer, format="JPEG")
img_bytes = buffer.getvalue()
# Convert to base64 and return as data URI
return f"data:image/jpeg;base64,{base64.b64encode(img_bytes).decode('utf-8')}"
except Exception as e:
print(f"Failed to create error image: {e}")
# Return a simple error message as fallback
return "Error: " + error_message
# Fix the poll_once function to accept a manager parameter
async def poll_once(manager, backend, request_id):
"""Poll once and return result if complete, otherwise None"""
headers = {"Authorization": f"Bearer {API_KEY}"}
url = f"{API_BASE_URL}/predictions/{request_id}/result"
async with aiohttp.ClientSession() as session:
async with session.get(url, headers=headers) as response:
if response.status == 200:
result = await response.json()
data = result["data"]
current_status = data["status"]
if current_status == "completed":
# IMPORTANT: Update status BEFORE returning - using the passed manager
manager.backend_statuses[backend].complete()
manager.update_activity()
# Handle base64 output
output = data["outputs"][0]
# Check if it's a base64 string or URL
if isinstance(output, str) and output.startswith("http"):
# It's a URL - return as is
return output
else:
# It's base64 data - format it as a data URI if needed
try:
# Format as data URI for Gradio to display directly
if isinstance(
output, str
) and not output.startswith("data:image"):
# Convert raw base64 to data URI format
return f"data:image/jpeg;base64,{output}"
else:
# Already in data URI format
return output
except Exception as e:
print(f"Error processing base64 image: {e}")
raise Exception(
f"Failed to process base64 image: {str(e)}")
elif current_status == "failed":
manager.backend_statuses[backend].fail()
manager.update_activity()
error = data.get("error", "Unknown error")
raise Exception(error)
# Still processing
return None
else:
raise Exception(f"Poll error: {response.status}")
# Store recent generations
recent_generations = []
# Example prompts
example_prompts = [
"A Ghibli-inspired scene featuring a giant, friendly creature walking through a lush forest and holding a sign saying 'HiDream', with vibrant colors, soft lighting, and whimsical details like floating lights and colorful, magical creatures. The scene captures a peaceful, enchanted atmosphere.",
"A neon-lit cyberpunk city at night, filled with towering skyscrapers, flying cars, and bustling streets. The air is thick with mist, and robots walk alongside humans, all illuminated by vibrant neon signs, holographic advertisements, and digital billboards.",
"A highly detailed, realistic portrait of an elderly man with deep wrinkles and a kind smile. Soft natural lighting, realistic skin textures, and lifelike details, capturing every strand of his white beard and the reflections in his glasses.",
]
# Use a state variable to store session ID
with gr.Blocks(theme=gr.themes.Soft()) as demo:
session_id = gr.State(None) # Add this to store session ID
# session_id = gr.BrowserState()(None) # Add this to store session ID
gr.Markdown("# 🌊 WaveSpeedAI HiDream Arena")
# Add the introduction with link to WaveSpeedAI
gr.Markdown("""
[WaveSpeedAI](https://wavespeed.ai/) is the global pioneer in accelerating AI-powered video and image generation.
Our in-house inference accelerator provides lossless speedup on image & video generation based on our rich inference optimization software stack, including our in-house inference compiler, CUDA kernel libraries and parallel computing libraries.
""")
gr.Markdown("""
This demo showcases the performance and outputs of leading image generation models, including HiDream and Flux, on our accelerated inference platform.
""")
with gr.Row():
with gr.Column(scale=3):
example_dropdown = gr.Dropdown(
choices=example_prompts,
label="Choose an example prompt",
interactive=True,
)
input_text = gr.Textbox(
example_prompts[0],
label="Enter your prompt",
placeholder="Type here...",
lines=3,
)
with gr.Column(scale=1):
generate_btn = gr.Button("Generate", variant="primary")
example_dropdown.change(lambda ex: ex,
inputs=[example_dropdown],
outputs=[input_text],
api_name=False)
# Two status boxes - small (default) and big (during generation)
small_status_box = gr.Markdown("Ready to generate images",
elem_id="small-status")
# Big status box in its own row with styling
with gr.Row(elem_id="big-status-row"):
big_status_box = gr.Markdown("",
elem_id="big-status",
visible=False,
elem_classes="big-status-box")
with gr.Row():
with gr.Column():
draft_output = gr.Image(label="Flux-dev")
with gr.Column():
quick_output = gr.Image(label="HiDream-dev")
with gr.Column():
best_output = gr.Image(label="HiDream-full")
performance_plot = gr.Plot(label="Performance Metrics")
with gr.Accordion("Recent Generations (last 16)", open=False):
recent_gallery = gr.Gallery(label="Prompt and Output",
columns=3,
interactive=False)
def get_recent_gallery_items():
gallery_items = []
for r in reversed(recent_generations):
if any(x is None for x in r.values()):
continue
gallery_items.append((r["flux-dev"], f"FLUX-dev: {r['prompt']}"))
gallery_items.append(
(r["hidream-dev"], f"HiDream-dev: {r['prompt']}"))
gallery_items.append(
(r["hidream-full"], f"HiDream-full: {r['prompt']}"))
return gallery_items
def update_recent_gallery(prompt, results):
recent_generations.append({
"prompt": prompt,
"flux-dev": results["flux-dev"],
"hidream-dev": results["hidream-dev"],
"hidream-full": results["hidream-full"],
})
if len(recent_generations) > 16:
recent_generations.pop(0)
gallery_items = get_recent_gallery_items()
return gr.update(value=gallery_items)
# Add custom CSS for the big status box
css = """
#big-status-row {
margin: 20px 0;
}
#big-status {
font-size: 28px; /* Even larger font size */
font-weight: bold;
padding: 30px; /* More padding */
background-color: #0D47A1; /* Deeper blue background */
color: white; /* White text */
border-radius: 10px;
text-align: center;
margin: 0 auto;
box-shadow: 0 6px 12px rgba(0, 0, 0, 0.2); /* Stronger shadow */
animation: deep-breath 3s infinite; /* Slower, deeper breathing animation */
width: 100%; /* Full width */
max-width: 800px; /* Maximum width */
transition: all 0.3s ease; /* Smooth transitions */
border-left: 6px solid #64B5F6; /* Add a colored border */
border-right: 6px solid #64B5F6; /* Add a colored border */
}
/* Deeper breathing animation */
@keyframes deep-breath {
0% {
opacity: 0.7;
transform: scale(0.98);
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.2);
}
50% {
opacity: 1;
transform: scale(1.01);
box-shadow: 0 8px 16px rgba(0, 0, 0, 0.3);
}
100% {
opacity: 0.7;
transform: scale(0.98);
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.2);
}
}
"""
gr.HTML(f"<style>{css}</style>")
# Update the generation function to use session manager
async def generate_all_backends_with_status_boxes(prompt,
current_session_id,
request: gr.Request):
client_ip = request.client.host
x_forwarded_for = request.headers.get('x-forwarded-for')
if x_forwarded_for:
client_ip = x_forwarded_for
print(f"Client IP: {client_ip}")
client_generation_manager = ClientManager.get_manager(client_ip)
client_generation_manager.update_activity()
if client_generation_manager.has_exceeded_limit(limit=20):
error_message = "❌ Your network has exceeded the limit of 20 requests per hour. Please try again later."
yield (
error_message,
error_message,
gr.update(visible=False),
gr.update(visible=True),
None,
None,
None,
None,
current_session_id, # Return the session ID
None,
)
return
client_generation_manager.add_request_timestamp()
"""Generate images with big status box during generation"""
# Get or create a session manager
session_id, manager = SessionManager.get_manager(current_session_id)
manager.update_activity()
# Check if the user has exceeded the request limit
if manager.has_exceeded_limit(
limit=10): # Set the limit to 10 requests per hour
error_message = "❌ You have exceeded the limit of 10 requests per hour. Please try again later."
yield (
error_message,
error_message,
gr.update(visible=False),
gr.update(visible=True),
None,
None,
None,
None,
session_id,
None,
)
return
# Add the current request timestamp
manager.add_request_timestamp()
# IMPORTANT: Reset history when starting a new generation
if prompt and prompt.strip() != "":
manager.reset_history() # Clear previous performance metrics
if not prompt or prompt.strip() == "":
# Handle empty prompt case
yield (
"⚠️ Please enter a prompt first",
"⚠️ Please enter a prompt first",
gr.update(visible=True),
gr.update(visible=False),
None,
None,
None,
None,
session_id, # Return the session ID
None,
)
return
# Check if the prompt is safe
classification, message = classify_text(prompt)
if classification != 0:
# Handle unsafe prompt case
yield (
message,
message,
gr.update(visible=True),
gr.update(visible=False),
None,
None,
None,
None,
session_id, # Return the session ID
None,
)
return
# Status message
status_message = f"🔄 PROCESSING: '{prompt}'"
# Initial state - clear all images, show big status box
yield (
status_message,
status_message,
gr.update(visible=True),
gr.update(visible=False),
None,
None,
None,
None,
session_id, # Return the session ID
None,
)
# For production mode:
completed_backends = set()
results = {"flux-dev": None, "hidream-dev": None, "hidream-full": None}
try:
# Submit all tasks
request_ids = {}
for backend in BACKENDS:
try:
request_id = await manager.submit_task(backend, prompt)
request_ids[backend] = request_id
except Exception as e:
# Handle submission error
print(f"Error submitting task for {backend}: {e}")
results[backend] = create_error_image(backend, str(e))
completed_backends.add(backend)
# Poll all backends until they complete
max_poll_attempts = 300
poll_attempt = 0
# Main polling loop
while len(completed_backends
) < 3 and poll_attempt < max_poll_attempts:
poll_attempt += 1
# Poll each pending backend
for backend in list(BACKENDS.keys()):
if backend in completed_backends:
continue
try:
# Only do actual API calls every few attempts to reduce load
if poll_attempt % 2 == 0 or backend == "flux-dev":
# Use the session manager instead of global manager
result = await poll_once(manager, backend,
request_ids[backend])
if result: # Backend completed
results[backend] = result
completed_backends.add(backend)
# Yield updated state when an image completes
yield (
status_message,
status_message,
gr.update(visible=True),
gr.update(visible=False),
results["flux-dev"],
results["hidream-dev"],
results["hidream-full"],
(manager.get_performance_plot()
if any(completed_backends) else None),
session_id,
None,
)
except Exception as e:
print(f"Error polling {backend}: {str(e)}")
# Wait between poll attempts
await asyncio.sleep(0.1)
# Final status
final_status = ("✅ All generations completed!"
if len(completed_backends) == 3 else
"⚠️ Some generations timed out")
gallery_update = update_recent_gallery(prompt, results)
# Final yield
yield (
final_status,
final_status,
gr.update(visible=False),
gr.update(visible=True),
results["flux-dev"],
results["hidream-dev"],
results["hidream-full"],
manager.get_performance_plot(),
session_id,
gallery_update,
)
except Exception as e:
# Error handling
error_message = f"❌ Error: {str(e)}"
yield (
error_message,
error_message,
gr.update(visible=False),
gr.update(visible=True),
None,
None,
None,
None,
session_id,
None,
)
# Schedule periodic cleanup of old sessions
def cleanup_task():
SessionManager.cleanup_old_sessions()
ClientManager.cleanup_old_clients()
# Schedule the next cleanup
threading.Timer(3600, cleanup_task).start() # Run every hour
# Start the cleanup task
cleanup_task()
# Update the click handler to include session_id
generate_btn.click(
fn=generate_all_backends_with_status_boxes,
inputs=[input_text, session_id],
outputs=[
small_status_box,
big_status_box,
big_status_box, # visibility
small_status_box, # visibility
draft_output,
quick_output,
best_output,
performance_plot,
session_id, # Update the session ID
recent_gallery, # Update the gallery
],
# api_name="generate",
api_name=False,
max_batch_size=10, # Process up to 10 requests at once
concurrency_limit=20, # Allow up to 20 concurrent requests
concurrency_id="generation", # Group concurrent requests under this ID
)
# Launch with increased max_threads
if __name__ == "__main__":
demo.queue(max_size=50).launch(
server_name="0.0.0.0",
max_threads=16, # Increase thread count for better concurrency
)
demo.queue(max_size=4).launch(
server_name="0.0.0.0",
max_threads=16, # Increase thread count for better concurrency
)