app.py

import spaces
import argparse
import os
import time
from os import path
import shutil
from datetime import datetime
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
import gradio as gr
import torch
from diffusers import FluxPipeline
from PIL import Image
from transformers import pipeline
import base64

translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")

# Hugging Face 토큰 설정
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN is None:
    raise ValueError("HF_TOKEN environment variable is not set")

# Setup and initialization code
cache_path = path.join(path.dirname(path.abspath(__file__)), "models")
PERSISTENT_DIR = os.environ.get("PERSISTENT_DIR", ".")
gallery_path = path.join(PERSISTENT_DIR, "gallery")

os.environ["TRANSFORMERS_CACHE"] = cache_path
os.environ["HF_HUB_CACHE"] = cache_path
os.environ["HF_HOME"] = cache_path

torch.backends.cuda.matmul.allow_tf32 = True

# Create gallery directory if it doesn't exist
if not path.exists(gallery_path):
    os.makedirs(gallery_path, exist_ok=True)

# 샘플 이미지와 프롬프트 정의
SAMPLE_IMAGES = {
    
    "3d2.webp": "the most famous hero according to Yuri Milner ",   
    "3d3.webp": "purple nest ",   
    "3d4.webp": "Timothy's sabbath ",   
    "3d5.webp": " A schoolboy friend of Julián Carax, fun-loving and loyal ",   
    "3d6.webp": "Friend of Daniel and his father ",   
    "3d7.webp": "WHERE ships of purple gently toss On seas of daffodil ",   
    "3d8.webp": "Beat the drums of tragedy for me, And let the white violins whir thin and slow ",   
    "3d9.webp": "And let the choir sing a stormy song To drown the rattle of my dying breath. ",   
    "3d10.webp": "Beat the drums of tragedy and death ",   
    "3d11.webp": "Beat the drums of tragedy for me. ",   
    "3d12.webp": "Touching the infinite, else far and untrod, With oracles divine that speak of God. ",   
    "3d13.webp": "Night, standing on her starry pulpit, free, Utters them in the dread, the silver roll Of spheres, woods, winds and waves, alternately ",   
    "3d14.webp": "On sermons deep, fit time to feast the soul. ",   
    "3d15.webp": "The bee is cradled in the bud; and far, Cold glittering lights, the azure curtain, throng— Planet on beaming planet, star on star. ",   
    "3d16.webp": "The lark's sweet pipe has ceased its latest song ",   
    "3d17.webp": "the snake was a roaming dog ",  
    "3d18.webp": "Antonio Battistella portraying Father of Giulia ",   
    "3d19.webp": "So straight to her father the brisk young lady went, And said, grant me one favour, do give your consent ",   
    "3d20.webp": "Before that we are marry’d let me your father see, All fear is, now" miscarry’d, my heart is full of glee ",   
    "3d21.webp": "My heart you now have gained, you are all I prize, So make yourself contented, pray be satisfied. ",   
    "3d22.webp": "O pray what is the favour that of me you crave? If it lies in my power you the same shall have ",   
    "3d23.webp": "Could I but see your father, and my mind reveal, I have both gold and silver, and houses at my will ",   

    "3d1.webp": "the most famous hero according to Zhou Qi"
}

class timer:
    def __init__(self, method_name="timed process"):
        self.method = method_name
    def __enter__(self):
        self.start = time.time()
        print(f"{self.method} starts")
    def __exit__(self, exc_type, exc_val, exc_tb):
        end = time.time()
        print(f"{self.method} took {str(round(end - self.start, 2))}s")

# Model initialization
if not path.exists(cache_path):
    os.makedirs(cache_path, exist_ok=True)

# 인증된 모델 로드
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16,
    use_auth_token=HF_TOKEN
)

# Hyper-SD LoRA 로드
pipe.load_lora_weights(
    hf_hub_download(
        "ByteDance/Hyper-SD",
        "Hyper-FLUX.1-dev-8steps-lora.safetensors",
        use_auth_token=HF_TOKEN
    )
)
pipe.fuse_lora(lora_scale=0.125)
pipe.to(device="cuda", dtype=torch.bfloat16)

def save_image(image):
    """Save the generated image and return the path"""
    try:
        if not os.path.exists(gallery_path):
            try:
                os.makedirs(gallery_path, exist_ok=True)
            except Exception as e:
                print(f"Failed to create gallery directory: {str(e)}")
                return None
        
        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
        random_suffix = os.urandom(4).hex()
        filename = f"generated_{timestamp}_{random_suffix}.png"
        filepath = os.path.join(gallery_path, filename)
        
        try:
            if isinstance(image, Image.Image):
                image.save(filepath, "PNG", quality=100)
            else:
                image = Image.fromarray(image)
                image.save(filepath, "PNG", quality=100)
            
            return filepath
        except Exception as e:
            print(f"Failed to save image: {str(e)}")
            return None
            
    except Exception as e:
        print(f"Error in save_image: {str(e)}")
        return None

def get_random_seed():
    return torch.randint(0, 1000000, (1,)).item()

@spaces.GPU
def process_and_save_image(height=1024, width=1024, steps=8, scales=3.5, prompt="", seed=None):
    global pipe
    
    if seed is None:
        seed = torch.randint(0, 1000000, (1,)).item()
    
    # 한글 감지 및 번역
    def contains_korean(text):
        return any(ord('가') <= ord(c) <= ord('힣') for c in text)
    
    # 프롬프트 전처리
    if contains_korean(prompt):
        translated = translator(prompt)[0]['translation_text']
        prompt = translated
    
    formatted_prompt = f"wbgmsst, 3D, {prompt}, white background"
    
    with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16), timer("inference"):
        try:
            generated_image = pipe(
                prompt=[formatted_prompt],
                generator=torch.Generator().manual_seed(int(seed)),
                num_inference_steps=int(steps),
                guidance_scale=float(scales),
                height=int(height),
                width=int(width),
                max_sequence_length=256
            ).images[0]
            
            saved_path = save_image(generated_image)
            if saved_path is None:
                print("Warning: Failed to save generated image")
            
            return generated_image
        except Exception as e:
            print(f"Error in image generation: {str(e)}")
            return None

def update_random_seed():
    """버튼으로 눌렀을 때 새로운 시드를 업데이트"""
    return gr.update(value=get_random_seed())

# Gradio 인터페이스
with gr.Blocks(
    theme=gr.themes.Soft(),
    css="""
        .container {
            background: linear-gradient(to bottom right, #1a1a1a, #4a4a4a);
            border-radius: 20px;
            padding: 20px;
        }
        .generate-btn {
            background: linear-gradient(45deg, #2196F3, #00BCD4);
            border: none;
            color: white;
            font-weight: bold;
            border-radius: 10px;
        }
        .output-image {
            border-radius: 15px;
            box-shadow: 0 8px 16px rgba(0,0,0,0.2);
        }
        .fixed-width {
            max-width: 1024px;
            margin: auto;
        }
        .gallery-container {
            margin-top: 40px;
            padding: 20px;
            background: #f5f5f5;
            border-radius: 15px;
        }
        .gallery-title {
            text-align: center;
            margin-bottom: 20px;
            color: #333;
            font-size: 1.5rem;
        }
    """
) as demo:
    gr.HTML(
        """
        <div style="text-align: center; max-width: 800px; margin: 0 auto; padding: 20px;">
            <h1 style="font-size: 2.5rem; color: #2196F3;">3D Style Image Generator</h1>
            <p style="font-size: 1.2rem; color: #666;">Create amazing 3D-style images with AI</p>
        </div>
        """
    )
    
    with gr.Row(elem_classes="container"):
        with gr.Column(scale=3):
            prompt = gr.Textbox(
                label="Image Description",
                placeholder="Describe the 3D image you want to create...",
                lines=3
            )
            
            with gr.Accordion("Advanced Settings", open=False):
                with gr.Row():
                    height = gr.Slider(
                        label="Height",
                        minimum=256,
                        maximum=1152,
                        step=64,
                        value=1024
                    )
                    width = gr.Slider(
                        label="Width",
                        minimum=256,
                        maximum=1152,
                        step=64,
                        value=1024
                    )
                
                with gr.Row():
                    steps = gr.Slider(
                        label="Inference Steps",
                        minimum=6,
                        maximum=25,
                        step=1,
                        value=8
                    )
                    scales = gr.Slider(
                        label="Guidance Scale",
                        minimum=0.0,
                        maximum=5.0,
                        step=0.1,
                        value=3.5
                    )
                
                seed = gr.Number(
                    label="Seed (random by default, set for reproducibility)",
                    value=get_random_seed(),
                    precision=0
                )
                
                randomize_seed = gr.Button("🎲 Randomize Seed", elem_classes=["generate-btn"])
            
            generate_btn = gr.Button(
                "✨ Generate Image",
                elem_classes=["generate-btn"]
            )

        with gr.Column(scale=4, elem_classes=["fixed-width"]):
            output = gr.Image(
                label="Generated Image",
                elem_id="output-image",
                elem_classes=["output-image", "fixed-width"],
                value="3d.webp"
            )

    # Gallery 섹션
    with gr.Row(elem_classes="gallery-container"):
        gr.HTML("<h2 class='gallery-title'>Gallery</h2>")
        gallery_html = "<div style='display: grid; grid-template-columns: repeat(auto-fill, minmax(300px, 1fr)); gap: 20px;'>"
        
        for img_file, prompt_text in SAMPLE_IMAGES.items():
            img_path = os.path.abspath(img_file)  # 절대 경로로 변환
            if os.path.exists(img_path):
                try:
                    # Base64로 이미지 인코딩
                    with open(img_path, "rb") as img:
                        img_data = base64.b64encode(img.read()).decode()
                    
                    # 3d1.webp 파일이므로 MIME타입은 webp
                    gallery_html += f"""
                    <div style='
                        border: 1px solid #ddd;
                        border-radius: 10px;
                        padding: 10px;
                        background: white;
                        box-shadow: 0 4px 8px rgba(0,0,0,0.1);
                    '>
                        <img src='data:image/webp;base64,{img_data}'
                            style='width: 100%; 
                            border-radius: 8px;
                            margin-bottom: 10px;'
                        >
                        <p style='
                            margin: 5px 0;
                            font-weight: bold;
                            color: #333;
                            padding: 10px;
                        '>Prompt: {prompt_text}</p>
                    </div>
                    """
                except Exception as e:
                    print(f"Error loading image {img_file}: {str(e)}")
        
        gallery_html += "</div>"
        gr.HTML(gallery_html)
    
    # 이벤트 핸들러
    generate_btn.click(
        fn=process_and_save_image,
        inputs=[height, width, steps, scales, prompt, seed],  # 딕셔너리 대신 리스트 형태
        outputs=output
    )

    randomize_seed.click(
        fn=update_random_seed,
        inputs=None,
        outputs=seed
    )

if __name__ == "__main__":
    demo.launch(allowed_paths=[PERSISTENT_DIR])