app.py

import os
import random
import torch
import gradio as gr
import numpy as np
import spaces
from diffusers import DiffusionPipeline
from PIL import Image

# --- [Optional Patch] ---------------------------------------------------------
# This patch fixes potential JSON schema parsing issues in Gradio/Gradio-Client.
import gradio_client.utils
original_json_schema = gradio_client.utils._json_schema_to_python_type

from PIL import ImageOps, ExifTags

def preprocess_image(image):
    # EXIF 정보에 따라 이미지 회전 조정
    try:
        image = ImageOps.exif_transpose(image)
    except Exception as e:
        print(f"EXIF 변환 오류: {e}")
    
    # 이미지 크기 조정 (너무 크면 모델이 제대로 처리하지 못할 수 있음)
    if max(image.width, image.height) > 1024:
        image.thumbnail((1024, 1024), Image.LANCZOS)
    
    # 이미지 모드 확인 및 변환
    if image.mode != "RGB":
        image = image.convert("RGB")
    
    return image

# DELETE THIS LINE COMPLETELY

def patched_json_schema(schema, defs=None):
    # Handle boolean schema directly
    if isinstance(schema, bool):
        return "bool"
    
    # If 'additionalProperties' is a boolean, replace it with a generic type
    try:
        if "additionalProperties" in schema and isinstance(schema["additionalProperties"], bool):
            schema["additionalProperties"] = {"type": "any"}
    except (TypeError, KeyError):
        pass
    
    # Attempt to parse normally; fallback to "any" on error
    try:
        return original_json_schema(schema, defs)
    except Exception:
        return "any"

gradio_client.utils._json_schema_to_python_type = patched_json_schema
# -----------------------------------------------------------------------------

# ----------------------------- Model Loading ----------------------------------
device = "cuda" if torch.cuda.is_available() else "cpu"
repo_id = "black-forest-labs/FLUX.1-dev"
adapter_id = "openfree/flux-chatgpt-ghibli-lora"

def load_model_with_retry(max_retries=5):
    for attempt in range(max_retries):
        try:
            print(f"Loading model attempt {attempt+1}/{max_retries}...")
            pipeline = DiffusionPipeline.from_pretrained(
                repo_id,
                torch_dtype=torch.bfloat16,
                use_safetensors=True,
                resume_download=True
            )
            print("Base model loaded successfully, now loading LoRA weights...")
            pipeline.load_lora_weights(adapter_id)
            pipeline = pipeline.to(device)
            print("Pipeline is ready!")
            return pipeline
        except Exception as e:
            if attempt < max_retries - 1:
                wait_time = 10 * (attempt + 1)
                print(f"Error loading model: {e}. Retrying in {wait_time} seconds...")
                import time
                time.sleep(wait_time)
            else:
                raise Exception(f"Failed to load model after {max_retries} attempts: {e}")

pipeline = load_model_with_retry()

# ----------------------------- Inference Function -----------------------------
MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024

@spaces.GPU(duration=120)
def inference(
    prompt: str,
    seed: int,
    randomize_seed: bool,
    width: int,
    height: int,
    guidance_scale: float,
    num_inference_steps: int,
    lora_scale: float,
):
    # If "randomize_seed" is selected, choose a random seed
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    
    print(f"Running inference with prompt: {prompt}")

    try:
        image = pipeline(
            prompt=prompt,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            width=width,
            height=height,
            generator=generator,
            joint_attention_kwargs={"scale": lora_scale},
        ).images[0]
        return image, seed
    except Exception as e:
        print(f"Error during inference: {e}")
        # Return a red error image of the specified size and the used seed
        error_img = Image.new('RGB', (width, height), color='red')
        return error_img, seed

# ----------------------------- Florence-2 Captioner ---------------------------
import subprocess
try:
    subprocess.run(
        'pip install flash-attn --no-build-isolation',
        env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"},
        shell=True
    )
except Exception as e:
    print(f"Warning: Could not install flash-attn: {e}")

from transformers import AutoProcessor, AutoModelForCausalLM

# Function to safely load models
def load_caption_model(model_name):
    try:
        model = AutoModelForCausalLM.from_pretrained(
            model_name, trust_remote_code=True
        ).eval()
        processor = AutoProcessor.from_pretrained(
            model_name, trust_remote_code=True
        )
        return model, processor
    except Exception as e:
        print(f"Error loading caption model {model_name}: {e}")
        return None, None

# Pre-load models and processors
print("Loading captioning models...")
default_caption_model = 'microsoft/Florence-2-large'
models = {}
processors = {}

# Try to load the default model
default_model, default_processor = load_caption_model(default_caption_model)
if default_model is not None and default_processor is not None:
    models[default_caption_model] = default_model
    processors[default_caption_model] = default_processor
    print(f"Successfully loaded default caption model: {default_caption_model}")
else:
    # Fallback to simpler model
    fallback_model = 'gokaygokay/Florence-2-Flux'
    fallback_model_obj, fallback_processor = load_caption_model(fallback_model)
    if fallback_model_obj is not None and fallback_processor is not None:
        models[fallback_model] = fallback_model_obj
        processors[fallback_model] = fallback_processor
        default_caption_model = fallback_model
        print(f"Loaded fallback caption model: {fallback_model}")
    else:
        print("WARNING: Failed to load any caption model!")

@spaces.GPU
def caption_image(image, model_name=default_caption_model):
    """
    Runs the selected Florence-2 model to generate a detailed caption.
    """
    from PIL import Image as PILImage
    import numpy as np
    
    print(f"Starting caption generation with model: {model_name}")
    
    # Handle case where image is already a PIL image
    if isinstance(image, PILImage.Image):
        pil_image = image
    else:
        # Convert numpy array to PIL
        if isinstance(image, np.ndarray):
            pil_image = PILImage.fromarray(image)
        else:
            print(f"Unexpected image type: {type(image)}")
            return "Error: Unsupported image type"
    
    # Convert input to RGB if needed
    if pil_image.mode != "RGB":
        pil_image = pil_image.convert("RGB")
    
    # Check if model is available
    if model_name not in models or model_name not in processors:
        available_models = list(models.keys())
        if available_models:
            model_name = available_models[0]
            print(f"Requested model not available, using: {model_name}")
        else:
            return "Error: No caption models available"
    
    model = models[model_name]
    processor = processors[model_name]
    
    task_prompt = "<DESCRIPTION>"
    user_prompt = task_prompt + "Describe this image in great detail."
    
    try:
        inputs = processor(text=user_prompt, images=pil_image, return_tensors="pt")
        
        generated_ids = model.generate(
            input_ids=inputs["input_ids"],
            pixel_values=inputs["pixel_values"],
            max_new_tokens=1024,
            num_beams=3,
            repetition_penalty=1.10,
        )
        
        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
        parsed_answer = processor.post_process_generation(
            generated_text, task=task_prompt, image_size=(pil_image.width, pil_image.height)
        )
        
        # Extract the caption
        caption = parsed_answer.get("<DESCRIPTION>", "")
        print(f"Generated caption: {caption}")
        return caption
    except Exception as e:
        print(f"Error during captioning: {e}")
        return f"Error generating caption: {str(e)}"

# --------- Process uploaded image and generate Ghibli style image ---------
@spaces.GPU(duration=120)
def process_uploaded_image(
    image,
    seed,
    randomize_seed,
    width,
    height,
    guidance_scale,
    num_inference_steps,
    lora_scale
):
    if image is None:
        print("No image provided")
        return None, None, "No image provided", "No image provided"
    
    print("Starting image processing workflow")
    
    # Step 1: Generate caption from the uploaded image
    try:
        caption = caption_image(image)
        if caption.startswith("Error:"):
            print(f"Captioning failed: {caption}")
            # Use a default caption as fallback
            caption = "A beautiful scene"
    except Exception as e:
        print(f"Exception during captioning: {e}")
        caption = "A beautiful scene"
    
    # Step 2: Append "ghibli style" to the caption
    ghibli_prompt = f"{caption}, ghibli style"
    print(f"Final prompt for Ghibli generation: {ghibli_prompt}")
    
    # Step 3: Generate Ghibli-style image based on the caption
    try:
        generated_image, used_seed = inference(
            prompt=ghibli_prompt,
            seed=seed,
            randomize_seed=randomize_seed,
            width=width,
            height=height,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            lora_scale=lora_scale
        )
        
        print(f"Image generation complete with seed: {used_seed}")
        return generated_image, used_seed, caption, ghibli_prompt
    except Exception as e:
        print(f"Error generating image: {e}")
        error_img = Image.new('RGB', (width, height), color='red')
        return error_img, seed, caption, ghibli_prompt

# Define Ghibli Studio Theme
ghibli_theme = gr.themes.Soft(
    primary_hue="indigo",
    secondary_hue="blue",
    neutral_hue="slate",
    font=[gr.themes.GoogleFont("Nunito"), "ui-sans-serif", "sans-serif"],
    radius_size=gr.themes.sizes.radius_sm,
).set(
    body_background_fill="#f0f9ff",
    body_background_fill_dark="#0f172a",
    button_primary_background_fill="#6366f1",
    button_primary_background_fill_hover="#4f46e5",
    button_primary_text_color="#ffffff",
    block_title_text_weight="600",
    block_border_width="1px",
    block_shadow="0 4px 6px -1px rgb(0 0 0 / 0.1), 0 2px 4px -2px rgb(0 0 0 / 0.1)",
)

# Custom CSS for enhanced visuals
custom_css = """
.gradio-container {
    max-width: 1200px !important;
}
.main-header {
    text-align: center;
    margin-bottom: 1rem;
    font-weight: 800;
    font-size: 2.5rem;
    background: linear-gradient(90deg, #4338ca, #3b82f6);
    -webkit-background-clip: text;
    -webkit-text-fill-color: transparent;
    padding: 0.5rem;
}
.tagline {
    text-align: center;
    font-size: 1.2rem;
    margin-bottom: 2rem;
    color: #4b5563;
}
.image-preview {
    border-radius: 12px;
    overflow: hidden;
    box-shadow: 0 10px 15px -3px rgb(0 0 0 / 0.1), 0 4px 6px -4px rgb(0 0 0 / 0.1);
}
.panel-box {
    border-radius: 12px;
    background-color: rgba(255, 255, 255, 0.8);
    padding: 1rem;
    box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1), 0 2px 4px -2px rgb(0 0 0 / 0.1);
}
.control-panel {
    padding: 1rem;
    border-radius: 12px;
    background-color: rgba(255, 255, 255, 0.9);
    margin-bottom: 1rem;
    border: 1px solid #e2e8f0;
}
.section-header {
    font-weight: 600;
    font-size: 1.1rem;
    margin-bottom: 0.5rem;
    color: #4338ca;
}
.transform-button {
    font-weight: 600 !important;
    margin-top: 1rem !important;
}
.footer {
    text-align: center;
    color: #6b7280;
    margin-top: 2rem;
    font-size: 0.9rem;
}
.output-panel {
    background: linear-gradient(135deg, #f0f9ff, #e0f2fe);
    border-radius: 12px;
    padding: 1rem;
    border: 1px solid #bfdbfe;
}
"""

# ----------------------------- Gradio UI --------------------------------------
with gr.Blocks(analytics_enabled=False, theme=ghibli_theme, css=custom_css) as demo:
    gr.HTML(
        """
        <div class="main-header">Open Ghibli Studio</div>
        <div class="tagline">Transform your photos into magical Ghibli-inspired artwork</div>
        """
    )
    
    # Background image for the app
    gr.HTML(
        """
        <style>
        body {
            background-image: url('https://i.imgur.com/LxPQPR1.jpg');
            background-size: cover;
            background-position: center;
            background-attachment: fixed;
            background-repeat: no-repeat;
            background-color: #f0f9ff;
        }
        @media (max-width: 768px) {
            body {
                background-size: contain;
            }
        }
        </style>
        """
    )

    with gr.Row(equal_height=True):
        with gr.Column(scale=1):
            with gr.Group(elem_classes="panel-box"):
                gr.HTML('<div class="section-header">Upload Image</div>')
                upload_img = gr.Image(
                    label="Drop your image here", 
                    type="pil",
                    elem_classes="image-preview",
                    height=400
                )
                
                with gr.Accordion("Advanced Settings", open=False):
                    with gr.Group(elem_classes="control-panel"):
                        gr.HTML('<div class="section-header">Generation Controls</div>')
                        with gr.Row():
                            img2img_seed = gr.Slider(
                                label="Seed",
                                minimum=0,
                                maximum=MAX_SEED,
                                step=1,
                                value=42,
                                info="Set a specific seed for reproducible results"
                            )
                            img2img_randomize_seed = gr.Checkbox(
                                label="Randomize Seed", 
                                value=True,
                                info="Enable to get different results each time"
                            )
                        
                        with gr.Group():
                            gr.HTML('<div class="section-header">Image Size</div>')
                            with gr.Row():
                                img2img_width = gr.Slider(
                                    label="Width",
                                    minimum=256,
                                    maximum=MAX_IMAGE_SIZE,
                                    step=32,
                                    value=1024,
                                    info="Image width in pixels"
                                )
                                img2img_height = gr.Slider(
                                    label="Height",
                                    minimum=256,
                                    maximum=MAX_IMAGE_SIZE,
                                    step=32,
                                    value=1024,
                                    info="Image height in pixels"
                                )
                        
                        with gr.Group():
                            gr.HTML('<div class="section-header">Generation Parameters</div>')
                            with gr.Row():
                                img2img_guidance_scale = gr.Slider(
                                    label="Guidance Scale",
                                    minimum=0.0,
                                    maximum=10.0,
                                    step=0.1,
                                    value=3.5,
                                    info="Higher values follow the prompt more closely"
                                )
                                img2img_steps = gr.Slider(
                                    label="Steps",
                                    minimum=1,
                                    maximum=50,
                                    step=1,
                                    value=30,
                                    info="More steps = more detailed but slower generation"
                                )
                            
                            img2img_lora_scale = gr.Slider(
                                label="Ghibli Style Strength",
                                minimum=0.0,
                                maximum=1.0,
                                step=0.05,
                                value=1.0,
                                info="Controls the intensity of the Ghibli style"
                            )
                
                transform_button = gr.Button("Transform to Ghibli Style", variant="primary", elem_classes="transform-button")
        
        with gr.Column(scale=1):
            with gr.Group(elem_classes="output-panel"):
                gr.HTML('<div class="section-header">Ghibli Magic Result</div>')
                ghibli_output_image = gr.Image(
                    label="Generated Ghibli Style Image", 
                    elem_classes="image-preview", 
                    height=400
                )
                ghibli_output_seed = gr.Number(label="Seed Used", interactive=False)
                
                # Debug elements
                with gr.Accordion("Image Details", open=False):
                    extracted_caption = gr.Textbox(
                        label="Detected Image Content",
                        placeholder="The AI will analyze your image and describe it here...",
                        info="AI-generated description of your uploaded image"
                    )
                    ghibli_prompt = gr.Textbox(
                        label="Generation Prompt",
                        placeholder="The prompt used to create your Ghibli image will appear here...",
                        info="Final prompt used for the Ghibli transformation"
                    )

    gr.HTML(
        """
        <div class="footer">
            <p>Open Ghibli Studio uses AI to transform your images into Ghibli-inspired artwork.</p>
            <p>Powered by FLUX.1 and Florence-2 models.</p>
        </div>
        """
    )

    # Auto-process when image is uploaded
    upload_img.upload(
        process_uploaded_image,
        inputs=[
            upload_img,
            img2img_seed,
            img2img_randomize_seed,
            img2img_width,
            img2img_height,
            img2img_guidance_scale,
            img2img_steps,
            img2img_lora_scale,
        ],
        outputs=[
            ghibli_output_image,
            ghibli_output_seed,
            extracted_caption,
            ghibli_prompt,
        ]
    )
    
    # Manual process button
    transform_button.click(
        process_uploaded_image,
        inputs=[
            upload_img,
            img2img_seed,
            img2img_randomize_seed,
            img2img_width,
            img2img_height,
            img2img_guidance_scale,
            img2img_steps,
            img2img_lora_scale,
        ],
        outputs=[
            ghibli_output_image,
            ghibli_output_seed,
            extracted_caption,
            ghibli_prompt,
        ]
    )

demo.launch(debug=True)