import torch from diffusers import StableDiffusionControlNetPipeline, ControlNetModel # <- KORREKT! from controlnet_aux import OpenposeDetector from PIL import Image import random import cv2 import numpy as np import gradio as gr class ControlNetProgressCallback: def __init__(self, progress, total_steps): self.progress = progress self.total_steps = total_steps self.current_step = 0 def __call__(self, pipe, step_index, timestep, callback_kwargs): self.current_step = step_index + 1 progress_percentage = self.current_step / self.total_steps if self.progress is not None: self.progress(progress_percentage, desc=f"ControlNet: Schritt {self.current_step}/{self.total_steps}") print(f"ControlNet Fortschritt: {self.current_step}/{self.total_steps} ({progress_percentage:.1%})") return callback_kwargs class ControlNetProcessor: def __init__(self, device="cuda", torch_dtype=torch.float32): self.device = device self.torch_dtype = torch_dtype self.pose_detector = None self.midas_model = None self.midas_transform = None def load_pose_detector(self): """Lädt nur den Pose-Detector""" if self.pose_detector is None: print("Loading Pose Detector...") try: self.pose_detector = OpenposeDetector.from_pretrained("lllyasviel/ControlNet") print("✅ Pose-Detector geladen") except Exception as e: print(f"⚠️ Pose-Detector konnte nicht geladen werden: {e}") return self.pose_detector def load_midas_model(self): """Lädt MiDaS Model für Depth Maps""" if self.midas_model is None: print("🔄 Lade MiDaS Modell für Depth Maps...") try: # WICHTIG: torchvision 0.20.0 hat MiDaS integriert import torchvision.transforms as T # MiDaS Small (weniger VRAM) self.midas_model = torch.hub.load( "intel-isl/MiDaS", "DPT_Hybrid", trust_repo=True ) self.midas_model.to(self.device) self.midas_model.eval() # Transform für MiDaS self.midas_transform = T.Compose([ T.Resize(384), T.ToTensor(), T.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]), ]) print("✅ MiDaS Modell erfolgreich geladen") except Exception as e: print(f"❌ MiDaS konnte nicht geladen werden: {e}") print("ℹ️ Verwende Fallback-Methode") self.midas_model = None return self.midas_model def extract_pose_simple(self, image): """Einfache Pose-Extraktion ohne komplexe Abhängigkeiten""" try: img_array = np.array(image.convert("RGB")) edges = cv2.Canny(img_array, 100, 200) pose_image = Image.fromarray(edges).convert("RGB") print("⚠️ Verwende Kanten-basierte Pose-Approximation") return pose_image except Exception as e: print(f"Fehler bei einfacher Pose-Extraktion: {e}") return image.convert("RGB").resize((512, 512)) def extract_pose(self, image): """Extrahiert Pose-Map aus Bild mit Fallback""" try: detector = self.load_pose_detector() if detector is None: return self.extract_pose_simple(image) pose_image = detector(image, hand_and_face=True) return pose_image except Exception as e: print(f"Fehler bei Pose-Extraktion: {e}") return self.extract_pose_simple(image) def extract_canny_edges(self, image): """Extrahiert Canny Edges für Umgebungserhaltung""" try: img_array = np.array(image.convert("RGB")) # Canny Edge Detection gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY) edges = cv2.Canny(gray, 100, 200) # Zu 3-Kanal Bild konvertieren edges_rgb = cv2.cvtColor(edges, cv2.COLOR_GRAY2RGB) edges_image = Image.fromarray(edges_rgb) print("✅ Canny Edge Map erstellt") return edges_image except Exception as e: print(f"Fehler bei Canny Edge Extraction: {e}") return image.convert("RGB").resize((512, 512)) def extract_depth_map(self, image): """ Extrahiert Depth Map mit MiDaS (Fallback auf Filter) """ try: # Versuche MiDaS midas = self.load_midas_model() if midas is not None: print("🎯 Verwende MiDaS für Depth Map...") import torchvision.transforms as T from PIL import Image # Bild vorbereiten img_transformed = self.midas_transform(image).unsqueeze(0).to(self.device) # Depth Map berechnen with torch.no_grad(): prediction = midas(img_transformed) prediction = torch.nn.functional.interpolate( prediction.unsqueeze(1), size=image.size[::-1], # (height, width) mode="bicubic", align_corners=False, ).squeeze() # Normalisieren für Ausgabe depth_np = prediction.cpu().numpy() depth_min, depth_max = depth_np.min(), depth_np.max() if depth_max > depth_min: depth_np = (depth_np - depth_min) / (depth_max - depth_min) depth_np = (depth_np * 255).astype(np.uint8) depth_image = Image.fromarray(depth_np).convert("RGB") print("✅ MiDaS Depth Map erfolgreich erstellt") return depth_image else: # Fallback auf einfache Methode print("⚠️ MiDaS nicht verfügbar, verwende Fallback...") raise Exception("MiDaS nicht geladen") except Exception as e: print(f"⚠️ MiDaS Fehler: {e}. Verwende Fallback...") # Fallback auf einfache Depth Map try: img_array = np.array(image.convert("RGB")) gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY) # Depth-ähnliche Map erstellen depth_map = cv2.GaussianBlur(gray, (5, 5), 0) depth_rgb = cv2.cvtColor(depth_map, cv2.COLOR_GRAY2RGB) depth_image = Image.fromarray(depth_rgb) print("✅ Fallback Depth Map erstellt") return depth_image except Exception as fallback_error: print(f"❌ Auch Fallback fehlgeschlagen: {fallback_error}") return image.convert("RGB").resize((512, 512)) def prepare_controlnet_maps(self, image, keep_environment=False): """ ERSTELLT NUR CONDITIONING-MAPS, generiert KEIN Bild. """ print("🎯 ControlNet: Erstelle Conditioning-Maps...") if keep_environment: # Depth + Canny print(" Modus: Depth + Canny") conditioning_images = [ self.extract_depth_map(image), self.extract_canny_edges(image) ] else: # OpenPose + Canny print(" Modus: OpenPose + Canny") conditioning_images = [ self.extract_pose(image), self.extract_canny_edges(image) ] print(f"✅ {len(conditioning_images)} Conditioning-Maps erstellt.") return conditioning_images # Rückgabe: Liste der PIL Images def prepare_inpaint_input(self, image, keep_environment=False): """ Bereitet das Input-Bild für Inpaint vor """ if keep_environment: print("🎯 Inpaint: Depth+Canny Info (Outside-Box ändern)") depth_image = self.extract_depth_map(image) canny_image = self.extract_canny_edges(image) combined_map = Image.blend(depth_image.convert("RGB"), canny_image.convert("RGB"), alpha=0.5) return combined_map, {"type": "depth_canny", "image": combined_map} else: print("🎯 Inpaint: Originalbild (Inside-Box ändern)") return image, {"type": "original", "image": image} # Globale Instanz device = "cuda" if torch.cuda.is_available() else "cpu" torch_dtype = torch.float16 if device == "cuda" else torch.float32 controlnet_processor = ControlNetProcessor(device=device, torch_dtype=torch_dtype)