Update app.py

app.py

@@ -56,7 +56,7 @@ def auto_negative_prompt(positive_prompt):
     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"
+    "girl", "boy", "lady", "gentleman", "fairy", "elf", "dwarf", "santa claus",
     "mermaid", "angel", "demon", "witch", "wizard", "creature", "being",
     "model", "actor", "actress", "celebrity", "avatar", "group"]):
         negatives.append(
@@ -336,9 +336,10 @@ def load_txt2img(model_id):
             raise
 
 def load_img2img():
+    """Lädt das Inpainting-Modell mit DPMSolver++ Scheduler"""
     global pipe_img2img
     if pipe_img2img is None:
-        print("Loading Inpainting model...")
+        print("🔄 Lade Inpainting-Modell...")
         try:
             pipe_img2img = StableDiffusionInpaintPipeline.from_pretrained(
                 "runwayml/stable-diffusion-inpainting",
@@ -346,22 +347,26 @@ def load_img2img():
                 allow_pickle=False,
                 safety_checker=None,
             ).to(device)
+            
+            # WICHTIG: Behalte DPMSolver++ Scheduler bei (beste Qualität für Inpainting)
+            pipe_img2img.scheduler = DPMSolverMultistepScheduler.from_config(
+                pipe_img2img.scheduler.config,
+                algorithm_type="sde-dpmsolver++",
+                use_karras_sigmas=True,
+                timestep_spacing="trailing"
+            )
+            
+            print("✅ DPMSolver++ Multistep Scheduler für Inpainting konfiguriert")
+            
         except Exception as e:
-            print(f"Fehler beim Laden des Inpainting-Modells: {e}")
+            print(f"❌ Fehler beim Laden des Inpainting-Modells: {e}")
             raise
 
-        from diffusers import DPMSolverMultistepScheduler
-        pipe_img2img.scheduler = DPMSolverMultistepScheduler.from_config(
-            pipe_img2img.scheduler.config,
-            algorithm_type="sde-dpmsolver++",
-            use_karras_sigmas=True,
-            timestep_spacing="trailing"
-        )
-        
         pipe_img2img.enable_attention_slicing()
         pipe_img2img.enable_vae_tiling()
         if hasattr(pipe_img2img, 'vae_slicing'):
             pipe_img2img.vae_slicing = True
+        print("✅ Inpainting-Modell geladen und optimiert")
 
     return pipe_img2img
 
@@ -767,47 +772,36 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
 
         pipe = load_img2img()
         
-        img_resized = inpaint_input.convert("RGB").resize((512, 512))
+        # ===== RICHTIGE BILD-SKALIERUNG =====
+        # Prüfe ob ControlNet schon 512×512 liefert
+        if inpaint_input.size != (512, 512):
+            img_resized = inpaint_input.convert("RGB").resize((512, 512), Image.Resampling.LANCZOS)
+            print("🔄 ControlNet Output von {} auf 512×512 skaliert".format(inpaint_input.size))
+        else:
+            img_resized = inpaint_input.convert("RGB")
+            print("✅ ControlNet Output ist bereits 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"🌱 Inpaint Seed: {seed}")
 
-        # ===== MASKE ERSTELLEN (BASIEREND AUF MODUS) =====
+        # ===== RICHTIGE MASKEN-ERSTELLUNG (FIX: KEINE DOPPELTE SKALIERUNG) =====
         mask = None
         if bbox_x1 is not None and bbox_y1 is not None and bbox_x2 is not None and bbox_y2 is not None:
-            # Skaliere Slider-Werte (Original-Bildgröße) auf 512x512 für die Pipeline
-            width, height = image.size
-            
-            # Skalierungsfaktoren berechnen
-            scale_x = IMG_SIZE / width
-            scale_y = IMG_SIZE / height
-            
-            # Skaliere Bounding-Box-Koordinaten auf 512x512
-            pipeline_x1 = int(bbox_x1 * scale_x)
-            pipeline_y1 = int(bbox_y1 * scale_y)
-            pipeline_x2 = int(bbox_x2 * scale_x)
-            pipeline_y2 = int(bbox_y2 * scale_y)
-            
-            # Sortiere und begrenze die Koordinaten
-            pipeline_bbox = sort_coordinates(
-                max(0, min(pipeline_x1, IMG_SIZE-1)),
-                max(0, min(pipeline_y1, IMG_SIZE-1)),
-                max(0, min(pipeline_x2, IMG_SIZE-1)),
-                max(0, min(pipeline_y2, IMG_SIZE-1))
-            )
+            # 1. Maske AUF ORIGINAL-BILD erstellen (korrekte Proportionen)
+            original_mask = create_face_mask(image, (bbox_x1, bbox_y1, bbox_x2, bbox_y2), mode)
             
-            print(f"Original BBox: [{bbox_x1}, {bbox_y1}, {bbox_x2}, {bbox_y2}] -> Pipeline BBox: {pipeline_bbox}")
+            # 2. Maske auf 512×512 skalieren (gleicher Algorithmus wie Bild)
+            # WICHTIG: NEAREST für scharfe Kanten, da Maske binär ist
+            mask = original_mask.resize((512, 512), Image.Resampling.NEAREST)
             
-            # Erstelle Maske basierend auf skalierten Koordinaten
-            mask = create_face_mask(img_resized, pipeline_bbox, mode)
-            if mask:
-                print(f"✅ Maske erstellt für Modus: {mode}")
+            print(f"✅ Maske erstellt für Modus: {mode}")
+            print(f"   Original BBox: [{bbox_x1}, {bbox_y1}, {bbox_x2}, {bbox_y2}]")
+            print(f"   Maske skaliert von {original_mask.size} auf {mask.size}")
 
-        from diffusers import EulerAncestralDiscreteScheduler
-        if not isinstance(pipe.scheduler, EulerAncestralDiscreteScheduler):
-            pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
+        # WICHTIG: KEINE SCHEDULER-ÄNDERUNG MEHR - DPMSolver++ bleibt aktiv
+        print(f"✅ Verwende DPMSolver++ Scheduler: {type(pipe.scheduler).__name__}")
 
         callback = ImageToImageProgressCallback(progress, int(steps), adj_strength)
 
@@ -830,6 +824,7 @@ def img_to_image(image, prompt, neg_prompt, strength, steps, guidance_scale,
         print(f"✅ Transformation abgeschlossen in {duration:.2f} Sekunden")
         print(f"🎯 Verwendeter Modus: {mode}")
         print(f"⚙️ ControlNet: {'Depth+Canny' if keep_environment else 'OpenPose+Canny'}")
+        print(f"⚙️ Scheduler: DPMSolver++ (optimal für Inpainting)")
 
         generated_image = result.images[0]
         return generated_image
@@ -856,7 +851,7 @@ def update_model_settings(model_id):
     return (
         config["recommended_steps"],  # steps
         config["recommended_cfg"],    # guidance_scale
-        f"📊 Empfohlene Einstellungen: {config['steps']} Steps, CFG {config['cfg']}"
+        f"📊 Empfohlene Einstellungen: {config['recommended_steps']} Steps, CFG {config['recommended_cfg']}"
     )
 
 def main_ui():