app.py

import os

# Set this environment variable to disable torch.compiler features
os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "1"
os.environ["TRANSFORMERS_COMPILER_DISABLED"] = "1"

import yaml
import torch
import sys
sys.path.append(os.path.abspath('./'))
from inference.utils import *
from train import WurstCoreB
from gdf import DDPMSampler
from train import WurstCore_t2i as WurstCoreC
import numpy as np
import random
import argparse
import gradio as gr
import spaces
from huggingface_hub import hf_hub_url
import subprocess
from huggingface_hub import hf_hub_download
from transformers import pipeline

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

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--height', type=int, default=2560, help='image height')
    parser.add_argument('--width', type=int, default=5120, help='image width')
    parser.add_argument('--seed', type=int, default=123, help='random seed')
    parser.add_argument('--dtype', type=str, default='bf16', help='if bf16 does not work, change it to float32')
    parser.add_argument('--config_c', type=str, 
    default='configs/training/t2i.yaml', help='config file for stage c, latent generation')
    parser.add_argument('--config_b', type=str, 
    default='configs/inference/stage_b_1b.yaml', help='config file for stage b, latent decoding')
    parser.add_argument('--prompt', type=str,
     default='A photo-realistic image of a west highland white terrier in the garden, high quality, detail rich, 8K', help='text prompt')
    parser.add_argument('--num_image', type=int, default=1, help='how many images generated')
    parser.add_argument('--output_dir', type=str, default='figures/output_results/', help='output directory for generated image')
    parser.add_argument('--stage_a_tiled', action='store_true', help='whether or not to use tiled decoding for stage a to save memory')
    parser.add_argument('--pretrained_path', type=str, default='models/ultrapixel_t2i.safetensors', help='pretrained path of newly added parameter of UltraPixel')
    args = parser.parse_args()
    return args

def clear_image():
    return None

def load_message(height, width, seed, prompt, args, stage_a_tiled):
    args.height = height
    args.width = width
    args.seed = seed
    args.prompt = prompt + ' rich detail, 4k, high quality'
    args.stage_a_tiled = stage_a_tiled
    return args

def is_korean(text):
    return any('\uac00' <= char <= '\ud7a3' for char in text)

def translate_if_korean(text, translation_display=None):
    if is_korean(text):
        translated = translator(text, max_length=512)[0]['translation_text']
        print(f"Translated from Korean: {text} -> {translated}")
        
        # If translation_display is provided, update it to show the translation
        if translation_display is not None:
            translation_display.update(value=f"Original (Korean): {text}\nTranslation (English): {translated}", visible=True)
            
        return translated
    
    # If not Korean, hide the translation display if it exists
    if translation_display is not None:
        translation_display.update(visible=False)
    
    return text

@spaces.GPU(duration=120)
def get_image(height, width, seed, prompt, cfg, timesteps, stage_a_tiled):
    global args
    
    args = load_message(height, width, seed, prompt, args, stage_a_tiled)
    torch.manual_seed(args.seed)
    random.seed(args.seed) 
    np.random.seed(args.seed)
    dtype = torch.bfloat16 if args.dtype == 'bf16' else torch.float

    captions = [args.prompt] * args.num_image
    height, width = args.height, args.width
    batch_size = 1 
    height_lr, width_lr = get_target_lr_size(height / width, std_size=32)
    stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
    stage_c_latent_shape_lr, stage_b_latent_shape_lr = calculate_latent_sizes(height_lr, width_lr, batch_size=batch_size)
   
    # Stage C Parameters
    extras.sampling_configs['cfg'] = 4
    extras.sampling_configs['shift'] = 1
    extras.sampling_configs['timesteps'] = 20
    extras.sampling_configs['t_start'] = 1.0
    extras.sampling_configs['sampler'] = DDPMSampler(extras.gdf)
    
    # Stage B Parameters
    extras_b.sampling_configs['cfg'] = 1.1
    extras_b.sampling_configs['shift'] = 1
    extras_b.sampling_configs['timesteps'] = 10
    extras_b.sampling_configs['t_start'] = 1.0

    for _, caption in enumerate(captions):
        batch = {'captions': [caption] * batch_size}
        conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
        unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
        
        with torch.no_grad():
            models.generator.cuda()
            print('STAGE C GENERATION***************************')
            with torch.cuda.amp.autocast(dtype=dtype):
                sampled_c = generation_c(batch, models, extras, core, stage_c_latent_shape, stage_c_latent_shape_lr, device)
            
            models.generator.cpu()
            torch.cuda.empty_cache()
            
            conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
            unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
            conditions_b['effnet'] = sampled_c
            unconditions_b['effnet'] = torch.zeros_like(sampled_c)
            print('STAGE B + A DECODING***************************')
            
            with torch.cuda.amp.autocast(dtype=dtype):
                sampled = decode_b(conditions_b, unconditions_b, models_b, stage_b_latent_shape, extras_b, device, stage_a_tiled=args.stage_a_tiled)
            
            torch.cuda.empty_cache()
            imgs = show_images(sampled)
                    
    return imgs[0]

def process_and_generate(height, width, seed, prompt, cfg, timesteps, stage_a_tiled, translation_display):
    # First translate if needed and show the translation
    translated_prompt = translate_if_korean(prompt, translation_display)
    
    # Then generate the image
    return get_image(height, width, seed, translated_prompt, cfg, timesteps, stage_a_tiled)

css = """
footer {
    visibility: hidden;
}

/* Main container styling */
#col-container {
    max-width: 1200px;
    margin: 0 auto;
    padding: 20px;
    background-color: #f8f9fa;
    border-radius: 15px;
    box-shadow: 0 4px 15px rgba(0, 0, 0, 0.1);
}

/* Header styling */
h1 {
    text-align: center;
    color: #ff6b00;
    font-size: 2.5rem;
    margin-bottom: 20px;
    font-weight: 700;
    text-shadow: 1px 1px 2px rgba(0,0,0,0.1);
}

/* Button styling */
button.primary {
    background-color: #ff6b00 !important;
    color: white !important;
    border: none !important;
    border-radius: 8px !important;
    padding: 10px 20px !important;
    font-weight: 600 !important;
    transition: all 0.3s ease !important;
}

button.primary:hover {
    background-color: #e55f00 !important;
    transform: translateY(-2px);
    box-shadow: 0 4px 8px rgba(0, 0, 0, 0.2) !important;
}

/* Input field styling */
input[type="text"] {
    border-radius: 8px !important;
    border: 2px solid #ddd !important;
    padding: 12px !important;
    font-size: 1rem !important;
    transition: all 0.3s ease !important;
}

input[type="text"]:focus {
    border-color: #ff6b00 !important;
    box-shadow: 0 0 0 3px rgba(255, 107, 0, 0.2) !important;
}

/* Output image container */
.output-image {
    border-radius: 12px;
    overflow: hidden;
    box-shadow: 0 8px 20px rgba(0, 0, 0, 0.15);
    margin: 20px 0;
    transition: all 0.3s ease;
}

.output-image:hover {
    transform: scale(1.02);
}

/* Accordion styling */
.accordion {
    border-radius: 10px !important;
    overflow: hidden !important;
    margin: 15px 0 !important;
    border: 1px solid #eaeaea !important;
}

/* Example gallery */
.examples-gallery {
    display: grid;
    grid-template-columns: repeat(auto-fill, minmax(300px, 1fr));
    gap: 15px;
    margin-top: 20px;
}

.example-item {
    background-color: white;
    border-radius: 10px;
    padding: 10px;
    box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1);
    transition: all 0.3s ease;
}

.example-item:hover {
    transform: translateY(-5px);
    box-shadow: 0 5px 15px rgba(0, 0, 0, 0.2);
}

/* Language indicator */
.language-indicator {
    background-color: #e6f7ff;
    color: #0072b5;
    border-radius: 20px;
    padding: 2px 10px;
    font-size: 0.8rem;
    margin-left: 10px;
    display: inline-block;
}

/* Loading animation */
@keyframes pulse {
    0% { opacity: 0.6; }
    50% { opacity: 1; }
    100% { opacity: 0.6; }
}

.loading {
    animation: pulse 1.5s infinite;
    text-align: center;
    padding: 20px;
    color: #ff6b00;
    font-weight: bold;
}
"""

with gr.Blocks(theme="Yntec/HaleyCH_Theme_Orange", css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("<h1>UHD Image Generator (5120×4096)</h1>")
        
        with gr.Group():
            with gr.Row():
                with gr.Column(scale=5):
                    prompt = gr.Textbox(
                        label="Text Prompt",
                        max_lines=2,
                        placeholder="Enter your prompt in Korean or English (한국어나 영어로 프롬프트를 입력하세요)",
                        elem_id="prompt-input"
                    )
                
                with gr.Column(scale=1):
                    generate_button = gr.Button("Generate Image", variant="primary", elem_id="generate-btn")
                    clear_button = gr.Button("Clear", elem_id="clear-btn")
        
        # Language detection indicator
        with gr.Row(visible=False) as language_indicator:
            gr.Markdown('<div class="language-indicator">Korean detected - will translate to English</div>')
        
        # Translation display component
        with gr.Row(visible=False) as translation_display:
            translation_text = gr.Textbox(
                label="Translation Information",
                interactive=False
            )
        
        # Loading indicator
        with gr.Row(visible=False) as loading_indicator:
            gr.Markdown('<div class="loading">Generating your ultra high resolution image... This may take a minute...</div>')
        
        # Output image with nicer styling
        output_img = gr.Image(label="Generated Image", elem_classes="output-image")
        
        with gr.Accordion("Advanced Settings", open=False):
            with gr.Row():
                with gr.Column():
                    seed = gr.Number(
                        label="Random Seed",
                        value=123,
                        step=1,
                        minimum=0,
                    )
                    
                    cfg = gr.Slider(
                        label="CFG Scale (Creativity vs. Prompt Adherence)",
                        minimum=3,
                        maximum=10,
                        step=0.1,
                        value=4
                    )
                
                with gr.Column():
                    with gr.Row():
                        width = gr.Slider(
                            label="Width",
                            minimum=1536,
                            maximum=5120,
                            step=32,
                            value=4096
                        )
                        
                        height = gr.Slider(
                            label="Height",
                            minimum=1536,
                            maximum=4096,
                            step=32,
                            value=2304
                        )
                    
                    timesteps = gr.Slider(
                        label="Timesteps (Quality vs. Speed)",
                        minimum=10,
                        maximum=50,
                        step=1,
                        value=20
                    )
                    
                    stage_a_tiled = gr.Checkbox(
                        label="Use Tiled Decoding (Lower Memory Usage)",
                        value=False
                    )
        
        # Aspect ratio presets
        with gr.Row():
            gr.Markdown("### Quick Aspect Ratio Presets")
            
        with gr.Row():
            preset_landscape = gr.Button("Landscape (16:9)", size="sm")
            preset_portrait = gr.Button("Portrait (9:16)", size="sm")
            preset_square = gr.Button("Square (1:1)", size="sm")
            preset_ultrawide = gr.Button("Ultrawide (21:9)", size="sm")
            
        # Examples with better organization
        gr.Markdown("### Example Prompts")
        with gr.Row():
            example_tabs = gr.Tabs([
                gr.TabItem("Nature", gr.Examples(
                    examples=[
                        "A detailed view of a blooming magnolia tree, with large, white flowers and dark green leaves, set against a clear blue sky.",
                        "눈 덮인 산맥의 장엄한 전경, 푸른 하늘을 배경으로 한 고요한 호수가 있는 모습",
                    ],
                    inputs=[prompt],
                    outputs=[output_img],
                )),
                gr.TabItem("Animals", gr.Examples(
                    examples=[
                        "스웨터를 입은 악어",
                        "골든 리트리버 강아지가 푸른 잔디밭에서 빨간 공을 쫓는 귀여운 모습",
                    ],
                    inputs=[prompt],
                    outputs=[output_img],
                )),
                gr.TabItem("Anime", gr.Examples(
                    examples=[
                        "A vibrant anime scene of a young girl with long, flowing pink hair, big sparkling blue eyes, and a school uniform, standing under a cherry blossom tree with petals falling around her.",
                    ],
                    inputs=[prompt],
                    outputs=[output_img],
                )),
                gr.TabItem("Architecture", gr.Examples(
                    examples=[
                        "A cozy, rustic log cabin nestled in a snow-covered forest, with smoke rising from the stone chimney, warm lights glowing from the windows, and a path of footprints leading to the front door.",
                    ],
                    inputs=[prompt],
                    outputs=[output_img],
                )),
            ])

        # Function to set aspect ratio presets
        def set_landscape():
            return 5120, 2880
            
        def set_portrait():
            return 2880, 4096
            
        def set_square():
            return 3584, 3584
            
        def set_ultrawide():
            return 5120, 2160
            
        # Connect buttons to functions
        preset_landscape.click(set_landscape, outputs=[width, height])
        preset_portrait.click(set_portrait, outputs=[width, height])
        preset_square.click(set_square, outputs=[width, height])
        preset_ultrawide.click(set_ultrawide, outputs=[width, height])
        
        # Function to show Korean detection indicator
        def check_korean(text):
            if any('\uac00' <= char <= '\ud7a3' for char in text):
                return gr.update(visible=True)
            return gr.update(visible=False)
            
        # Connect events
        prompt.change(check_korean, inputs=[prompt], outputs=[language_indicator])
        generate_button.click(
            lambda: gr.update(visible=True),
            outputs=[loading_indicator]
        ).then(
            process_and_generate,
            inputs=[height, width, seed, prompt, cfg, timesteps, stage_a_tiled, translation_display],
            outputs=[output_img]
        ).then(
            lambda: gr.update(visible=False),
            outputs=[loading_indicator]
        )
        
        clear_button.click(clear_image, inputs=[], outputs=[output_img])

def download_with_wget(url, save_path):
    try:
        subprocess.run(['wget', url, '-O', save_path], check=True)
        print(f"Downloaded to {save_path}")
    except subprocess.CalledProcessError as e:
        print(f"Error downloading file: {e}")

def download_model():
    urls = [
        'https://huggingface.co/stabilityai/StableWurst/resolve/main/stage_a.safetensors',
        'https://huggingface.co/stabilityai/StableWurst/resolve/main/previewer.safetensors',
        'https://huggingface.co/stabilityai/StableWurst/resolve/main/effnet_encoder.safetensors',
        'https://huggingface.co/stabilityai/StableWurst/resolve/main/stage_b_lite_bf16.safetensors', 
        'https://huggingface.co/stabilityai/StableWurst/resolve/main/stage_c_bf16.safetensors', 
    ]
    for file_url in urls:
        hf_hub_download(repo_id="stabilityai/stable-cascade", filename=file_url.split('/')[-1], local_dir='models')
    hf_hub_download(repo_id="roubaofeipi/UltraPixel", filename='ultrapixel_t2i.safetensors', local_dir='models')
    
if __name__ == "__main__":
    args = parse_args()
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    download_model()
    config_file = args.config_c
    with open(config_file, "r", encoding="utf-8") as file:
        loaded_config = yaml.safe_load(file)
    
    core = WurstCoreC(config_dict=loaded_config, device=device, training=False)
    
    # SETUP STAGE B
    config_file_b = args.config_b
    with open(config_file_b, "r", encoding="utf-8") as file:
        config_file_b = yaml.safe_load(file)
        
    core_b = WurstCoreB(config_dict=config_file_b, device=device, training=False)
    
    extras = core.setup_extras_pre()
    models = core.setup_models(extras)
    models.generator.eval().requires_grad_(False)
    print("STAGE C READY")
    
    extras_b = core_b.setup_extras_pre()
    models_b = core_b.setup_models(extras_b, skip_clip=True)
    models_b = WurstCoreB.Models(
       **{**models_b.to_dict(), 'tokenizer': models.tokenizer, 'text_model': models.text_model}
    )
    models_b.generator.bfloat16().eval().requires_grad_(False)
    print("STAGE B READY")
    
    pretrained_path = args.pretrained_path    
    sdd = torch.load(pretrained_path, map_location='cpu')
    collect_sd = {}
    for k, v in sdd.items():
        collect_sd[k[7:]] = v
    
    models.train_norm.load_state_dict(collect_sd)
    models.generator.eval()
    models.train_norm.eval()
    
    demo.launch(debug=True, share=True)