app.py

import gradio as gr
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
from diffusers import StableDiffusionInpaintPipeline, AutoencoderKL
from diffusers import DPMSolverMultistepScheduler, PNDMScheduler
from controlnet_module import controlnet_processor
import torch
from PIL import Image, ImageDraw
import time
import os
import tempfile
import random
import re

# === FACE-FIX IMPORT (automatisch nur bei Personen) ===
try:
    from controlnet_facefix import apply_facefix
    FACEFIX_AVAILABLE = True
    print("Face-Fix (OpenPose_faceonly + Depth) erfolgreich geladen")
except Exception as e:
    print(f"Face-Fix nicht verfügbar: {e}")
    FACEFIX_AVAILABLE = False

# === OPTIMIERTE EINSTELLUNGEN ===
device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if device == "cuda" else torch.float32
IMG_SIZE = 512

print(f"Running on: {device}")

# === MODELLKONFIGURATION (NUR 2 MODELLE) ===
MODEL_CONFIGS = {
    "runwayml/stable-diffusion-v1-5": {
        "name": "Stable Diffusion 1.5 (Universal)",
        "description": "Universal model, good all-rounder, reliable results",
        "requires_vae": False,
        "recommended_steps": 35,
        "recommended_cfg": 7.5,
        "supports_fp16": True
    },
    "SG161222/Realistic_Vision_V6.0_B1_noVAE": {
        "name": "Realistic Vision V6.0 (Portraits)",
        "description": "Best for photorealistic faces, skin details, human portraits",
        "requires_vae": True,
        "vae_model": "stabilityai/sd-vae-ft-mse",
        "recommended_steps": 40,
        "recommended_cfg": 7.0,
        "supports_fp16": False
    }
}

SAFETENSORS_MODELS = ["runwayml/stable-diffusion-v1-5"]
current_model_id = "runwayml/stable-diffusion-v1-5"

# === AUTOMATISCHE NEGATIVE PROMPT GENERIERUNG ===
def auto_negative_prompt(positive_prompt):
    p = positive_prompt.lower()
    negatives = []

    if any(w in p for w in [
        "person", "man", "woman", "face", "portrait", "team", "employee", 
        "people", "crowd", "character", "figure", "human", "child", "baby",
        "girl", "boy", "lady", "gentleman", "fairy", "elf", "dwarf", "santa claus",
        "mermaid", "angel", "demon", "witch", "wizard", "creature", "being",
        "model", "actor", "actress", "celebrity", "avatar", "group"
    ]):
        negatives.append(
            "blurry face, lowres face, deformed pupils, bad anatomy, malformed hands, extra fingers, uneven eyes, distorted face, "
            "unrealistic skin, mutated, ugly, disfigured, poorly drawn face, "
            "missing limbs, extra limbs, fused fingers, too many fingers, bad teeth, "
            "mutated hands, long neck, extra wings, multiple wings, grainy face, noisy face, "
            "compression artifacts, rendering artifacts, digital artifacts, overprocessed face, oversmoothed face "
        )

    if any(w in p for w in ["office", "business", "team", "meeting", "corporate", "company", "workplace"]):
        negatives.append("overexposed, oversaturated, harsh lighting, watermark, text, logo, brand")

    if any(w in p for w in ["product", "packshot", "mockup", "render", "3d", "cgi", "packaging"]):
        negatives.append("plastic texture, noisy, overly reflective surfaces, watermark, text, low poly")

    if any(w in p for w in ["landscape", "nature", "mountain", "forest", "outdoor", "beach", "sky"]):
        negatives.append("blurry, oversaturated, unnatural colors, distorted horizon, floating objects")

    if any(w in p for w in ["logo", "symbol", "icon", "typography", "badge", "emblem"]):
        negatives.append("watermark, signature, username, text, writing, scribble, messy")

    if any(w in p for w in ["building", "architecture", "house", "interior", "room", "facade"]):
        negatives.append("deformed, distorted perspective, floating objects, collapsing structure")

    base_negatives = "low quality, worst quality, blurry, jpeg artifacts, ugly, deformed"
    
    return base_negatives + ", " + ", ".join(negatives) if negatives else base_negatives

# === PERSONEN-ERKENNUNG (für Face-Fix) ===
def is_person_prompt(prompt: str) -> bool:
    p = prompt.lower()
    person_keywords = [
        "person", "man", "woman", "face", "portrait", "people", "child", "girl", "boy",
        "fairy", "elf", "witch", "santa", "nikolaus", "human", "character", "figure"
    ]
    return any(w in p for w in person_keywords)

# === GESICHTSMASKEN-FUNKTIONEN ===
def create_face_mask(image, bbox_coords, face_preserve):
    mask = Image.new("L", image.size, 0)
    if bbox_coords and all(coord is not None for coord in bbox_coords):
        x1, y1, x2, y2 = bbox_coords
        draw = ImageDraw.Draw(mask)
        if face_preserve:
            draw.rectangle([0, 0, image.size[0], image.size[1]], fill=255)
            draw.rectangle([x1, y1, x2, y2], fill=0)
        else:
            draw.rectangle([x1, y1, x2, y2], fill=255)
    return mask

def auto_detect_face_area(image):
    width, height = image.size
    face_size = min(width, height) * 0.4
    x1 = (width - face_size) / 2
    y1 = (height - face_size) / 4
    x2 = x1 + face_size
    y2 = y1 + face_size * 1.2
    x1, y1 = max(0, int(x1)), max(0, int(y1))
    x2, y2 = min(width, int(x2)), min(height, int(y2))
    return [x1, y1, x2, y2]

# === PIPELINES ===
pipe_txt2img = None
current_pipe_model_id = None
pipe_img2img = None

def load_txt2img(model_id):
    global pipe_txt2img, current_pipe_model_id
    if pipe_txt2img is not None and current_pipe_model_id == model_id:
        return pipe_txt2img
    
    print(f"Lade Modell: {model_id}")
    config = MODEL_CONFIGS.get(model_id, MODEL_CONFIGS["runwayml/stable-diffusion-v1-5"])
    
    try:
        vae = None
        if config.get("requires_vae", False):
            vae = AutoencoderKL.from_pretrained(config["vae_model"], torch_dtype=torch_dtype).to(device)
        
        model_params = {
            "torch_dtype": torch_dtype,
            "safety_checker": None,
            "requires_safety_checker": False,
        }
        
        if model_id in SAFETENSORS_MODELS:
            model_params["use_safetensors"] = True
        
        if config.get("supports_fp16", False) and torch_dtype == torch.float16:
            model_params["variant"] = "fp16"
        
        if vae is not None:
            model_params["vae"] = vae
        
        pipe_txt2img = StableDiffusionPipeline.from_pretrained(model_id, **model_params).to(device)
        pipe_txt2img.enable_attention_slicing()
        
        try:
            pipe_txt2img.scheduler = DPMSolverMultistepScheduler.from_config(
                pipe_txt2img.scheduler.config,
                use_karras_sigmas=True,
                algorithm_type="sde-dpmsolver++"
            )
        except:
            pass
        
        current_pipe_model_id = model_id
        return pipe_txt2img
        
    except Exception as e:
        print(f"Fehler beim Laden, Fallback auf SD 1.5: {e}")
        pipe_txt2img = StableDiffusionPipeline.from_pretrained(
            "runwayml/stable-diffusion-v1-5", torch_dtype=torch_dtype, use_safetensors=True
        ).to(device)
        pipe_txt2img.enable_attention_slicing()
        current_pipe_model_id = "runwayml/stable-diffusion-v1-5"
        return pipe_txt2img

def load_img2img():
    global pipe_img2img
    if pipe_img2img is None:
        pipe_img2img = StableDiffusionInpaintPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting", torch_dtype=torch_dtype, safety_checker=None
        ).to(device)
        pipe_img2img.enable_attention_slicing()
        pipe_img2img.enable_vae_tiling()
    return pipe_img2img

# === CALLBACKS ===
class TextToImageProgressCallback:
    def __init__(self, progress, total_steps):
        self.progress = progress
        self.total_steps = total_steps
    def __call__(self, pipe, step, timestep, callback_kwargs):
        self.progress(step / self.total_steps, desc="Generierung läuft...")
        return callback_kwargs

class ImageToImageProgressCallback:
    def __init__(self, progress, total_steps, strength):
        self.progress = progress
        self.total_steps = total_steps
        self.strength = strength
        self.actual_steps = None
    def __call__(self, pipe, step, timestep, callback_kwargs):
        if self.actual_steps is None:
            self.actual_steps = int(self.total_steps * self.strength)
        progress_val = step / self.actual_steps
        self.progress(progress_val, desc="Generierung läuft...")
        return callback_kwargs

# === HAUPTFUNKTION: TEXT ZU BILD MIT AUTOMATISCHEM FACE-FIX ===
def text_to_image(prompt, model_id, steps, guidance_scale, progress=gr.Progress()):
    try:
        if not prompt or not prompt.strip():
            return None, "Bitte einen Prompt eingeben"
            
        print(f"Generierung mit Modell: {model_id}")
        auto_negatives = auto_negative_prompt(prompt)
        start_time = time.time()

        # Qualitäts-Boost nur wenn nicht vorhanden
        quality_keywords = ['masterpiece', 'best quality', 'raw', 'highly detailed', 'ultra realistic']
        has_quality = any(kw in prompt.lower() for kw in quality_keywords)
        has_weights = bool(re.search(r':\d+\.\d+|\([^)]+:\d', prompt))

        enhanced_prompt = f"masterpiece, raw, best quality, highly detailed, {prompt}" if not (has_quality or has_weights) else prompt

        progress(0, desc="Lade Modell...")
        pipe = load_txt2img(model_id)
        
        seed = random.randint(0, 2**32 - 1)
        generator = torch.Generator(device=device).manual_seed(seed)

        image = pipe(
            prompt=enhanced_prompt,
            negative_prompt=auto_negatives,
            height=512, width=512,
            num_inference_steps=int(steps),
            guidance_scale=guidance_scale,
            generator=generator,
            callback_on_step_end=TextToImageProgressCallback(progress, steps),
            callback_on_step_end_tensor_inputs=[],
        ).images[0]

        # AUTOMATISCHER FACE-FIX NUR BEI PERSONEN
        if FACEFIX_AVAILABLE and is_person_prompt(enhanced_prompt):
            print("Person erkannt → Starte 20-Sekunden Face-Fix...")
            progress(0.92, desc="Perfektioniere Gesicht & Hände...")
            try:
                image = apply_facefix(
                    image=image,
                    prompt=enhanced_prompt,
                    negative_prompt=auto_negatives,
                    seed=seed,
                    model_id=model_id
                )
                print("Face-Fix abgeschlossen!")
            except Exception as e:
                print(f"Face-Fix fehlgeschlagen (ignoriert): {e}")

        duration = time.time() - start_time
        config = MODEL_CONFIGS.get(model_id, {"name": model_id})
        status_msg = f"Generiert mit {config.get('name', model_id)} in {duration:.1f}s"
        return image, status_msg
        
    except Exception as e:
        print(f"Fehler in text_to_image: {e}")
        import traceback
        traceback.print_exc()
        return None, f"Fehler: {str(e)}"


def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
                 face_preserve, bbox_x1, bbox_y1, bbox_x2, bbox_y2,
                 progress=gr.Progress()):
    try:
        if image is None:
            return None

        import time, random
        start_time = time.time()

        print(f"Img2Img Start → Strength: {strength}, Steps: {steps}, Guidance: {guidance_scale}")
        print(f"Prompt: {prompt}")
        print(f"Negativ-Prompt: {neg_prompt}")
        print(f"Gesicht beibehalten: {face_preserve}")
        

                 # ===== NEU: AUTOMATISCHEN NEGATIV-PROMPT GENERIEREN =====
        auto_negatives = auto_negative_prompt(prompt)
        print(f"🤖 Automatisch generierter Negativ-Prompt: {auto_negatives}")
        
        # ===== NEU: KOMBINIERE MANUELLEN UND AUTOMATISCHEN PROMPT =====
        combined_negative_prompt = ""
        
        if neg_prompt and neg_prompt.strip():
            # Benutzer hat einen Negativ-Prompt eingegeben
            user_neg = neg_prompt.strip()
            print(f"👤 Benutzer Negativ-Prompt: {user_neg}")
            
            # Entferne Duplikate zwischen automatischen und manuellen Prompts
            # Konvertiere beide in Sets für einfachen Duplikatvergleich
            user_words = [word.strip().lower() for word in user_neg.split(",")]
            auto_words = [word.strip().lower() for word in auto_negatives.split(",")]
            
            # Starte mit dem Benutzer-Prompt
            combined_words = user_words.copy()
            
            # Füge automatische Wörter hinzu, die nicht bereits vorhanden sind
            for auto_word in auto_words:
                if auto_word and auto_word not in user_words:
                    combined_words.append(auto_word)
            
            # Zusammenfügen und Duplikate entfernen (für den Fall von Duplikaten innerhalb des gleichen Prompts)
            unique_words = []
            seen_words = set()
            for word in combined_words:
                if word and word not in seen_words:
                    unique_words.append(word)
                    seen_words.add(word)
            
            combined_negative_prompt = ", ".join(unique_words)
        else:
            # Kein Benutzer-Prompt, verwende nur den automatischen
            combined_negative_prompt = auto_negatives
            print(f"ℹ️ Kein manueller Negativ-Prompt, verwende nur automatischen: {combined_negative_prompt}")
        
        print(f"✅ Finaler kombinierter Negativ-Prompt: {combined_negative_prompt}")
        # ===== ENDE DER NEUEN LOGIK =====
        

        progress(0, desc="Starte Generierung mit ControlNet...")

        adj_strength = min(0.85, strength * 1.25)
        
        if face_preserve:
            controlnet_strength = adj_strength * 0.8
            print(f"🎯 ControlNet Modus: Umgebung beibehalten (Strength = {controlnet_strength:.3f})")
        else:
            controlnet_strength = adj_strength * 0.5
            print(f"🎯 ControlNet Modus: Person beibehalten (Strength = {controlnet_strength:.3f})")
            
        controlnet_steps = min(25, int(steps * 0.8))

        print(f"🎯 Steps={steps}, ControlNet-Steps={controlnet_steps}, Strength={controlnet_strength:.3f}")

        progress(0.05, desc="Erstelle ControlNet Maps...")

        controlnet_output, inpaint_input = controlnet_processor.generate_with_controlnet(
            image=image,
            prompt=prompt,
            negative_prompt=combined_negative_prompt,
            steps=controlnet_steps,
            guidance_scale=guidance_scale,
            controlnet_strength=controlnet_strength,
            progress=progress,
            keep_environment=face_preserve
        )

        print(f"✅ ControlNet Output erhalten: {type(controlnet_output)}")
        print(f"✅ Inpaint Input erhalten: {type(inpaint_input)}")

        progress(0.3, desc="ControlNet abgeschlossen – starte Inpaint...")

        pipe = load_img2img()
        
        img_resized = inpaint_input.convert("RGB").resize((512, 512))

        adj_guidance = min(guidance_scale, 12.0)
        seed = random.randint(0, 2**32 - 1)
        generator = torch.Generator(device=device).manual_seed(seed)
        print(f"Using seed: {seed}")

        mask = None
        if bbox_x1 and bbox_y1 and bbox_x2 and bbox_y2:
            orig_w, orig_h = image.size
            scale_x, scale_y = 512 / orig_w, 512 / orig_h
            bbox_coords = [
                int(bbox_x1 * scale_x),
                int(bbox_y1 * scale_y),
                int(bbox_x2 * scale_x),
                int(bbox_y2 * scale_y)
            ]
            print(f"Skalierte Koordinaten: {bbox_coords}")
            mask = create_face_mask(img_resized, bbox_coords, face_preserve)
            if mask:
                print("✅ Maske erfolgreich erstellt")
        else:
            print("⚠️ Keine gültigen Koordinaten – keine Maske")

        from diffusers import EulerAncestralDiscreteScheduler
        if not isinstance(pipe.scheduler, EulerAncestralDiscreteScheduler):
            pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

        callback = ImageToImageProgressCallback(progress, int(steps), adj_strength)

        result = pipe(
            prompt=prompt,
            negative_prompt=combined_negative_prompt,
            image=img_resized,
            mask_image=mask,
            strength=adj_strength,
            num_inference_steps=int(steps),
            guidance_scale=adj_guidance,
            generator=generator,
            callback_on_step_end=callback,
            callback_on_step_end_tensor_inputs=[],
        )

        end_time = time.time()
        print(f"🕒 Dauer: {end_time - start_time:.2f} Sekunden")

        generated_image = result.images[0]
        return generated_image

    except Exception as e:
        print(f"❌ Fehler in img_to_image: {e}")
        import traceback
        traceback.print_exc()
        return None

def update_bbox_from_image(image):
    """Aktualisiert die Bounding-Box-Koordinaten wenn ein Bild hochgeladen wird"""
    if image is None:
        return None, None, None, None
    
    bbox = auto_detect_face_area(image)
    return bbox[0], bbox[1], bbox[2], bbox[3]

def update_model_settings(model_id):
    """Aktualisiert die empfohlenen Einstellungen basierend auf Modellauswahl"""
    config = MODEL_CONFIGS.get(model_id, MODEL_CONFIGS["runwayml/stable-diffusion-v1-5"])
    
    return (
        config["recommended_steps"],  # steps
        config["recommended_cfg"],    # guidance_scale
        f"📊 Empfohlene Einstellungen: {config['steps']} Steps, CFG {config['cfg']}"
    )

def main_ui():
    with gr.Blocks(
        title="AI Image Generator", 
        theme=gr.themes.Base(),
        css="""
        .info-box {
            background-color: #f8f4f0; 
            padding: 15px; 
            border-radius: 8px; 
            border-left: 4px solid #8B7355;
            margin: 20px 0;
        }
        .clickable-file {
           color: #1976d2;
           cursor: pointer;
           text-decoration: none;
           font-family: 'Monaco', 'Consolas', monospace;
           background: #e3f2fd;
           padding: 2px 6px;
           border-radius: 4px;
           border: 1px solid #bbdefb;
         }
        .clickable-file:hover {
           background: #bbdefb;
           text-decoration: underline;
         }
        .model-info-box {
            background: #e8f4fd;
            padding: 12px;
            border-radius: 6px;
            margin: 10px 0;
            border-left: 4px solid #2196f3;
            font-size: 14px;
        }
        #generate-button {
            background-color: #0080FF !important;
            border: none !important;
            margin: 20px auto !important;
            display: block !important;
            font-weight: 600;
            width: 280px;
        }
        #generate-button:hover {
            background-color: #0066CC !important;
        }
        .hint-box {
            margin-top: 20px;
        }
        .custom-text {
            font-size: 25px !important;
        }
        .image-upload .svelte-1p4f8co {
            display: block !important;
        }
        .preview-box {
            border: 2px dashed #ccc;
            padding: 10px;
            border-radius: 8px;
            margin: 10px 0;
        }
        .mode-red {
            border: 3px solid #ff4444 !important;
        }
        .mode-green {
            border: 3px solid #44ff44 !important;
        }
        .coordinate-sliders {
            background: #f8f9fa;
            padding: 15px;
            border-radius: 8px;
            margin: 10px 0;
        }
        .gr-checkbox .wrap .text-gray {
            font-size: 14px !important;
            font-weight: 600 !important;
            line-height: 1.4 !important;
        }
        .status-message {
            padding: 10px;
            border-radius: 5px;
            margin: 10px 0;
            text-align: center;
            font-weight: 500;
        }
        .status-success {
            background-color: #d4edda;
            color: #155724;
            border: 1px solid #c3e6cb;
        }
        .status-error {
            background-color: #f8d7da;
            color: #721c24;
            border: 1px solid #f5c6cb;
        }
        """
    ) as demo:

        with gr.Column(visible=True) as content_area:
            with gr.Tab("Text zu Bild"):
                gr.Markdown("## 🎨 Text zu Bild Generator")
                
                with gr.Row():
                    with gr.Column(scale=2):
                        # Modellauswahl Dropdown (NUR 2 MODELLE)
                        model_dropdown = gr.Dropdown(
                            choices=[
                                (config["name"], model_id) 
                                for model_id, config in MODEL_CONFIGS.items()
                            ],
                            value="runwayml/stable-diffusion-v1-5",
                            label="📁 Modellauswahl",
                            info="🏠 Universal vs 👤 Portraits"
                        )
                        
                        # Modellinformationen Box
                        model_info_box = gr.Markdown(
                            value="<div class='model-info-box'>"
                                  "**🏠 Stable Diffusion 1.5 (Universal)**<br>"
                                  "Universal model, good all-rounder, reliable results<br>"
                                  "Empfohlene Einstellungen: 35 Steps, CFG 7.5"
                                  "</div>",
                            label="Modellinformationen"
                        )
                        
                    with gr.Column(scale=3):
                        txt_input = gr.Textbox(
                            placeholder="z.B. ultra realistic mountain landscape at sunrise, soft mist over the valley, detailed foliage, crisp textures, depth of field, sunlight rays through clouds, shot on medium format camera, 8k, HDR, hyper-detailed, natural lighting, masterpiece",
                            lines=3,
                            label="🎯 Prompt (Englisch)",
                            info="Beschreibe detailliert, was du sehen möchtest. Negative Prompts werden automatisch generiert."
                        )
                
                with gr.Row():
                    with gr.Column():
                        txt_steps = gr.Slider(
                            minimum=10, maximum=100, value=35, step=1,
                            label="⚙️ Inferenz-Schritte",
                            info="Mehr Schritte = bessere Qualität, aber langsamer (20-50 empfohlen)"
                        )
                    with gr.Column():
                        txt_guidance = gr.Slider(
                            minimum=1.0, maximum=20.0, value=7.5, step=0.5,
                            label="🎛️ Prompt-Stärke (CFG Scale)",
                            info="Wie stark der Prompt befolgt wird (7-12 für gute Balance)"
                        )
                
                # Status-Nachricht
                status_output = gr.Markdown(
                    value="",
                    elem_classes="status-message"
                )
                
                generate_btn = gr.Button("🚀 Bild generieren", variant="primary", elem_id="generate-button")
                
                with gr.Row():
                    txt_output = gr.Image(
                        label="🖼️ Generiertes Bild", 
                        show_download_button=True,
                        type="pil",
                        height=400
                    )
                
                # Event-Handler für Modelländerung
                def update_model_info(model_id):
                    config = MODEL_CONFIGS.get(model_id, MODEL_CONFIGS["runwayml/stable-diffusion-v1-5"])
                    info_html = f"""
                    <div class='model-info-box'>
                    <strong>{config['name']}</strong><br>
                    {config['description']}<br>
                    <em>Empfohlene Einstellungen: {config['recommended_steps']} Steps, CFG {config['recommended_cfg']}</em>
                    </div>
                    """
                    return info_html, config["recommended_steps"], config["recommended_cfg"]
                
                model_dropdown.change(
                    fn=update_model_info,
                    inputs=[model_dropdown],
                    outputs=[model_info_box, txt_steps, txt_guidance]
                )
                
                generate_btn.click(
                    fn=text_to_image,
                    inputs=[txt_input, model_dropdown, txt_steps, txt_guidance],
                    outputs=[txt_output, status_output],
                    concurrency_limit=1
                )

            with gr.Tab("Bild zu Bild"):
                gr.Markdown("## 🖼️ Bild zu Bild Transformation")
                
                with gr.Row():
                    with gr.Column():
                        img_input = gr.Image(
                            type="pil", 
                            label="📤 Eingabebild",
                            height=300,
                            sources=["upload"],
                            elem_id="image-upload"
                        )
                    with gr.Column():
                        preview_output = gr.Image(
                            label="🎯 Live-Vorschau mit Maske",
                            height=300,
                            interactive=False,
                            show_download_button=False
                        )
                
                with gr.Row():
                    face_preserve = gr.Checkbox(
                        label="🛡️ Schutzmodus", 
                        value=True,
                        info="🟢 AN: Alles AUSSERHALB des gelben Rahmens verändern | 🔴 AUS: Nur INNERHALB des gelben Rahmens verändern"
                    )
                
                with gr.Row():
                    gr.Markdown("### 📐 Bildelementbereich anpassen")
                
                with gr.Row():
                    with gr.Column():
                        bbox_x1 = gr.Slider(
                            label="← Links (x1)", 
                            minimum=0, maximum=512, value=100, step=1,
                            info="Linke Kante des Bildelementbereichs"
                        )
                    with gr.Column():
                        bbox_y1 = gr.Slider(
                            label="↑ Oben (y1)", 
                            minimum=0, maximum=512, value=100, step=1,
                            info="Obere Kante des Bildelementbereichs"
                        )
                with gr.Row():
                    with gr.Column():
                        bbox_x2 = gr.Slider(
                            label="→ Rechts (x2)", 
                            minimum=0, maximum=512, value=300, step=1,
                            info="Rechte Kante des Bildelementbereichs"
                        )
                    with gr.Column():
                        bbox_y2 = gr.Slider(
                            label="↓ Unten (y2)", 
                            minimum=0, maximum=512, value=300, step=1,
                            info="Untere Kante des Bildelementbereichs"
                        )
                
                with gr.Row():
                    with gr.Column():
                        img_prompt = gr.Textbox(
                            placeholder="change background to beach with palm trees, keep person unchanged, sunny day",
                            lines=2,
                            label="🎯 Transformations-Prompt (Englisch)",
                            info="Was soll verändert werden? Sei spezifisch."
                        )
                    with gr.Column():
                        img_neg_prompt = gr.Textbox(
                            placeholder="blurry, deformed, ugly, bad anatomy, extra limbs, poorly drawn hands",
                            lines=2,
                            label="🚫 Negativ-Prompt (Englisch)",
                            info="Was soll vermieden werden? Unerwünschte Elemente auflisten."
                        )
                
                with gr.Row():
                    with gr.Column():
                        strength_slider = gr.Slider(
                            minimum=0.1, maximum=0.9, value=0.4, step=0.05,
                            label="💪 Veränderungs-Stärke",
                            info="0.1-0.3: Leichte Anpassungen, 0.4-0.6: Mittlere Veränderungen, 0.7-0.9: Starke Umgestaltung"
                        )
                    with gr.Column():
                        img_steps = gr.Slider(
                            minimum=10, maximum=100, value=35, step=1,
                            label="⚙️ Inferenz-Schritte",
                            info="Anzahl der Verarbeitungsschritte (25-45 für gute Ergebnisse)"
                        )
                    with gr.Column():
                        img_guidance = gr.Slider(
                            minimum=1.0, maximum=20.0, value=7.5, step=0.5,
                            label="🎛️ Prompt-Stärke",
                            info="Einfluss des Prompts auf das Ergebnis (6-10 für natürliche Ergebnisse)"
                        )

                with gr.Row():
                   gr.Markdown(
                     "### 📋 Hinweise:\n"
                     "• **🆕 Automatische Bildelementerkennung** setzt Koordinaten beim Upload\n"  
                     "• **🆕 Live-Vorschau** zeigt farbige Rahmen je nach Modus (🔴 Rot / 🟢 Grün)\n"
                     "• **🆕 Koordinaten-Schieberegler** für präzise Anpassung mit Live-Update\n"
                     "• **Koordinaten nur bei erkennbaren Verzerrungen anpassen** (Bereiche leicht verschieben)"
                    )
                
                transform_btn = gr.Button("🔄 Bild transformieren", variant="primary")
                
                with gr.Row():
                    img_output = gr.Image(
                        label="✨ Transformiertes Bild",
                        show_download_button=True,
                        type="pil",
                        height=400
                    )
                
                img_input.change(
                    fn=process_image_upload,
                    inputs=[img_input],
                    outputs=[preview_output, bbox_x1, bbox_y1, bbox_x2, bbox_y2]
                )
                
                coordinate_inputs = [img_input, bbox_x1, bbox_y1, bbox_x2, bbox_y2, face_preserve]
                
                for slider in [bbox_x1, bbox_y1, bbox_x2, bbox_y2]:
                    slider.change(
                        fn=update_live_preview,
                        inputs=coordinate_inputs,
                        outputs=preview_output
                    )
                
                face_preserve.change(
                    fn=update_live_preview,
                    inputs=coordinate_inputs,
                    outputs=preview_output
                )
                
                transform_btn.click(
                    fn=img_to_image,
                    inputs=[
                        img_input, img_prompt, img_neg_prompt, 
                        strength_slider, img_steps, img_guidance, 
                        face_preserve, bbox_x1, bbox_y1, bbox_x2, bbox_y2
                    ],
                    outputs=img_output,
                    concurrency_limit=1
                )

    return demo

if __name__ == "__main__":
    demo = main_ui()
    demo.queue(max_size=3)
    demo.launch(
        server_name="0.0.0.0", 
        server_port=7860,
        max_file_size="10MB",
        show_error=True,
        share=False,
        ssr_mode=False  # SSR deaktivieren für Stabilität
    )