app.py

import spaces
import gradio as gr
import numpy as np
import PIL.Image
from PIL import Image
import random
from diffusers import StableDiffusionXLPipeline
from diffusers import EulerAncestralDiscreteScheduler
import torch
from compel import Compel, ReturnedEmbeddingsType
import gc
import os

# Check if CUDA is available
print(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"CUDA device: {torch.cuda.get_device_name(0)}")

# Initialize the pipeline ONCE at startup
print("Loading pipeline...")
pipe = StableDiffusionXLPipeline.from_pretrained(
    "dhead/wai-nsfw-illustrious-sdxl-v140-sdxl",
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True,
    low_cpu_mem_usage=True
)

pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

# Enable memory efficient attention if available
if hasattr(pipe, "enable_model_cpu_offload"):
    pipe.enable_model_cpu_offload()
elif hasattr(pipe, "enable_attention_slicing"):
    pipe.enable_attention_slicing()

print("Pipeline loaded successfully!")

# Initialize Compel for long prompt processing
compel = None

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1216

# Default prompt
DEFAULT_PROMPT = "Detailed illustration, realistic style, portrait of a beautiful Japanese woman, wearing an elegant traditional Japanese uniform, neatly tailored with intricate patterns and subtle textures, serene expression, soft natural lighting, standing gracefully in a traditional Japanese garden with cherry blossom petals gently falling in the background, cinematic quality, ultra-detailed, high-resolution, warm tones"

def initialize_compel():
    """Initialize Compel with the current pipeline's components"""
    global compel
    if compel is None:
        try:
            compel = Compel(
                tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
                text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
                returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
                requires_pooled=[False, True],
                truncate_long_prompts=False
            )
        except Exception as e:
            print(f"Failed to initialize Compel: {e}")
            compel = None
    return compel

def process_long_prompt(prompt, negative_prompt=""):
    """Simple long prompt processing using Compel"""
    try:
        comp = initialize_compel()
        if comp is None:
            return None, None
        conditioning, pooled = comp([prompt, negative_prompt])
        return conditioning, pooled
    except Exception as e:
        print(f"Long prompt processing failed: {e}, falling back to standard processing")
        return None, None

@spaces.GPU(duration=90)  # Increased duration for stability
def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps):
    try:
        # Move pipeline to GPU inside the GPU-decorated function
        pipe.to("cuda")
        
        # Ensure all components are on GPU with correct dtype
        pipe.text_encoder = pipe.text_encoder.to(dtype=torch.float16)
        pipe.text_encoder_2 = pipe.text_encoder_2.to(dtype=torch.float16)
        pipe.vae = pipe.vae.to(dtype=torch.float16)
        pipe.unet = pipe.unet.to(dtype=torch.float16)
        
        use_long_prompt = len(prompt.split()) > 60 or len(prompt) > 300
            
        if randomize_seed:
            seed = random.randint(0, MAX_SEED)

        generator = torch.Generator(device="cuda").manual_seed(seed)
        
        # Try long prompt processing first if prompt is long
        if use_long_prompt:
            print("Using long prompt processing...")
            conditioning, pooled = process_long_prompt(prompt, negative_prompt)
            
            if conditioning is not None:
                output_image = pipe(
                    prompt_embeds=conditioning[0:1],
                    pooled_prompt_embeds=pooled[0:1],
                    negative_prompt_embeds=conditioning[1:2],
                    negative_pooled_prompt_embeds=pooled[1:2],
                    guidance_scale=guidance_scale,
                    num_inference_steps=num_inference_steps,
                    width=width,
                    height=height,
                    generator=generator
                ).images[0]
                
                # Clear GPU cache
                torch.cuda.empty_cache()
                gc.collect()
                
                return output_image
        
        # Fall back to standard processing
        output_image = pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            width=width,
            height=height,
            generator=generator
        ).images[0]
        
        # Clear GPU cache
        torch.cuda.empty_cache()
        gc.collect()
        
        return output_image
        
    except RuntimeError as e:
        print(f"Runtime error during generation: {e}")
        torch.cuda.empty_cache()
        gc.collect()
        # Return a blank image with error message
        error_img = Image.new('RGB', (width, height), color=(50, 50, 50))
        return error_img
    except Exception as e:
        print(f"Unexpected error: {e}")
        torch.cuda.empty_cache()
        gc.collect()
        error_img = Image.new('RGB', (width, height), color=(100, 0, 0))
        return error_img

css = """
/* Main container styling */
#col-container {
    margin: 0 auto;
    max-width: 1024px;
}
/* Gradient background for the entire app */
.gradio-container {
    background: linear-gradient(135deg, #667eea 0%, #764ba2 25%, #f093fb 50%, #f5576c 75%, #ffc947 100%);
    min-height: 100vh;
}
/* Main block styling with semi-transparent background */
.contain {
    background: rgba(255, 255, 255, 0.95);
    border-radius: 20px;
    padding: 20px;
    box-shadow: 0 8px 32px 0 rgba(31, 38, 135, 0.37);
    backdrop-filter: blur(4px);
    border: 1px solid rgba(255, 255, 255, 0.18);
}
/* Input field styling */
.gr-text-input {
    background: rgba(255, 255, 255, 0.9) !important;
    border: 2px solid rgba(102, 126, 234, 0.3) !important;
    border-radius: 10px !important;
}
/* Button styling */
.gr-button {
    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
    border: none !important;
    color: white !important;
    font-weight: bold !important;
    transition: all 0.3s ease !important;
}
.gr-button:hover {
    transform: translateY(-2px);
    box-shadow: 0 5px 15px rgba(102, 126, 234, 0.4);
}
/* Accordion styling */
.gr-accordion {
    background: rgba(255, 255, 255, 0.8) !important;
    border-radius: 10px !important;
    margin-top: 10px !important;
}
/* Result image container */
.gr-image {
    border-radius: 15px !important;
    box-shadow: 0 4px 15px rgba(0, 0, 0, 0.1) !important;
}
/* Slider styling */
.gr-slider {
    background: rgba(255, 255, 255, 0.8) !important;
}
/* Additional styling for headers */
h1, h2, h3 {
    color: #333 !important;
    text-align: center;
}
/* Markdown text styling */
.gr-markdown {
    text-align: center;
    margin-bottom: 20px;
}
"""

print("Building Gradio interface...")

# Build the Gradio interface
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(
            """
            # 🎨 Stable Diffusion XL Image Generator
            ### Create stunning AI-generated images with advanced controls
            """
        )

        # Badge section
        gr.HTML(
            """
            <div style="display: flex; justify-content: center; align-items: center; gap: 20px; margin: 20px 0;">
                <a href="https://huggingface.co/spaces/Heartsync/Wan-2.2-ADULT" target="_blank">
                    <img src="https://img.shields.io/static/v1?label=T2I%20%26%20TI2V&message=Wan-2.2-ADULT&color=%230000ff&labelColor=%23800080&logo=huggingface&logoColor=white&style=for-the-badge" alt="badge">
                </a>
                <a href="https://huggingface.co/spaces/Heartsync/PornHUB" target="_blank">
                    <img src="https://img.shields.io/static/v1?label=T2I%20&message=PornHUB&color=%230000ff&labelColor=%23800080&logo=huggingface&logoColor=white&style=for-the-badge" alt="badge">
                </a>
                <a href="https://huggingface.co/spaces/Heartsync/Hentai-Adult" target="_blank">
                    <img src="https://img.shields.io/static/v1?label=T2I%20&message=Hentai-Adult&color=%230000ff&labelColor=%23800080&logo=huggingface&logoColor=white&style=for-the-badge" alt="badge">
                </a>
            </div>
            """
        )
        
        with gr.Row():
            prompt = gr.Text(
                label="Prompt",
                show_label=False,
                max_lines=1,
                placeholder="Enter your prompt (long prompts are automatically supported)",
                container=False,
                value=DEFAULT_PROMPT
            )

            run_button = gr.Button("Run", scale=0)

        result = gr.Image(format="png", label="Result", show_label=False)
        
        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Text(
                label="Negative prompt",
                max_lines=1,
                placeholder="Enter a negative prompt",
                value="monochrome, (low quality, worst quality:1.2), very displeasing, 3d, watermark, signature, ugly, poorly drawn,"
            )

            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )

                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=MAX_IMAGE_SIZE,
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=20.0,
                    step=0.1,
                    value=7,
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=28,
                    step=1,
                    value=28,
                )

    # Connect the run button to the inference function
    run_button.click(
        fn=infer,
        inputs=[prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
        outputs=[result]
    )

print("Starting Gradio app...")

# Launch the app - CRITICAL: This must be at the module level for Spaces
demo.queue(max_size=20)
demo.launch()