controlnet_facefix.py

# controlnet_facefix.py - KORRIGIERTE VERSION FÜR HF SPACES
import torch
from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel
from PIL import Image
import time

print("="*60)
print("CONTROLNET_FACEFIX START")
print("="*60)

# WICHTIG: Keine .to("cuda") beim Laden!
_components_loaded = False
_depth_processor = None
_controlnet_depth = None  
_controlnet_pose = None
_pipeline = None

def _initialize_components():
    """Lade alle Komponenten mit Fehlerbehandlung"""
    global _components_loaded, _depth_processor, _controlnet_depth, _controlnet_pose
    
    if _components_loaded:
        return True
        
    try:
        print("1. Lade Depth Processor...")
        from controlnet_aux import ZoeDetector
        _depth_processor = ZoeDetector.from_pretrained("lllyasviel/ControlNet")
        print("   ✅ Depth Processor OK")
    except Exception as e:
        print(f"   ❌ Depth Processor Fehler: {e}")
        return False
    
    try:
        print("2. Lade ControlNet Depth...")
        _controlnet_depth = ControlNetModel.from_pretrained(
            "lllyasviel/control_v11f1e_sd15_depth",
            torch_dtype=torch.float16  # KEIN .to("cuda") hier!
        )
        print("   ✅ ControlNet Depth OK")
    except Exception as e:
        print(f"   ❌ ControlNet Depth Fehler: {e}")
        return False
    
    try:
        print("3. Lade ControlNet OpenPose...")
        # WICHTIG: KEIN subfolder="faceonly" - das gibt es nicht!
        _controlnet_pose = ControlNetModel.from_pretrained(
            "lllyasviel/control_v11p_sd15_openpose",
            torch_dtype=torch.float16  # KEIN .to("cuda") hier!
        )
        print("   ✅ ControlNet OpenPose OK")
    except Exception as e:
        print(f"   ❌ ControlNet OpenPose Fehler: {e}")
        return False
    
    _components_loaded = True
    print("✅ ALLE KOMPONENTEN GELADEN")
    return True

def _create_control_images(image):
    """Erstelle Depth und Pose Bilder"""
    try:
        print("   Erstelle Depth Map...")
        depth_img = _depth_processor(image)
        
        print("   Erstelle OpenPose...")
        from controlnet_aux import OpenposeDetector
        openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
        pose_img = openpose(image)
        
        return pose_img, depth_img
    except Exception as e:
        print(f"   ❌ Control Images Fehler: {e}")
        return None, None

def apply_facefix(image: Image.Image, prompt: str, negative_prompt: str, seed: int, model_id: str):
    """HAUPTFUNKTION - Optimiert für HF Spaces"""
    print("\n" + "🎭"*50)
    print("FACE-FIX WIRD AUSGEFÜHRT")
    print(f"  Model: {model_id}")
    print(f"  Seed: {seed}")
    print("🎭"*50)
    
    start_time = time.time()
    
    # 1. Komponenten initialisieren
    if not _initialize_components():
        print("❌ Komponenten konnten nicht geladen werden")
        return image
    
    # 2. Control Images erstellen
    pose_img, depth_img = _create_control_images(image)
    if pose_img is None or depth_img is None:
        print("❌ Control Images konnten nicht erstellt werden")
        return image
    
    # 3. Pipeline erstellen (lazy)
    global _pipeline
    if _pipeline is None:
        try:
            print("🔄 Lade Pipeline...")
            _pipeline = StableDiffusionControlNetInpaintPipeline.from_pretrained(
                model_id,
                controlnet=[_controlnet_pose, _controlnet_depth],
                torch_dtype=torch.float16,
                safety_checker=None,
                requires_safety_checker=False,
            )
            
            # HF Spaces Optimierungen
            _pipeline.enable_attention_slicing()
            _pipeline.enable_vae_slicing()
            
            print("✅ Pipeline geladen")
        except Exception as e:
            print(f"❌ Pipeline Fehler: {e}")
            return image
    
    try:
        # 4. Auf GPU bewegen (erst JETZT!)
        device = "cuda" if torch.cuda.is_available() else "cpu"
        print(f"  Device: {device}")
        pipeline = _pipeline.to(device)
        
        # 5. Prompts optimieren
        face_prompt = f"{prompt}, perfect face, detailed skin, realistic eyes, sharp focus"
        face_negative = f"{negative_prompt}, deformed face, blurry face, bad anatomy, ugly"
        
        print("⚡ Führe Face-Fix Inference aus...")
        
        # 6. Face-Fix ausführen
        result = pipeline(
            prompt=face_prompt,
            negative_prompt=face_negative,
            image=image,
            mask_image=None,  # Ganzer Bildbereich
            control_image=[pose_img, depth_img],
            controlnet_conditioning_scale=[0.8, 0.6],  # OpenPose stärker
            strength=0.4,  # Mittlere Stärke
            num_inference_steps=20,
            guidance_scale=7.0,
            generator=torch.Generator(device).manual_seed(seed),
            height=512,
            width=512,
        ).images[0]
        
        duration = time.time() - start_time
        print(f"\n✅✅✅ FACE-FIX ERFOLGREICH in {duration:.1f}s ✅✅✅")
        
        return result
        
    except Exception as e:
        print(f"\n❌❌❌ FACE-FIX FEHLGESCHLAGEN: {e} ❌❌❌")
        import traceback
        traceback.print_exc()
        return image

print("="*60)
print("CONTROLNET_FACEFIX INITIALISIERT")
print(f"apply_facefix Funktion: {'apply_facefix' in globals()}")
print("="*60)