Update controlnet_facefix.py

controlnet_facefix.py

@@ -3,17 +3,13 @@ from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel
 from controlnet_aux import OpenposeDetector, ZoeDetector
 from PIL import Image
 
-# ───── Globale Modelle (einmal laden, bleibt im VRAM) ─────
-print("Lade OpenPose_faceonly + Depth für perfekte Gesichter/Hände...")
+print("Lade OpenPose_faceonly + Depth für perfekte Gesichter...")
 
-# OpenPose Face-Only Preprocessor + ControlNet
-openpose_face = OpenposeDetector.from_pretrained(
-    "lllyasviel/ControlNet", model_name="openpose_face"
-)
-
-# Depth Preprocessor + ControlNet
+# Preprocessors (einmalig laden)
+openpose_face = OpenposeDetector.from_pretrained("lllyasviel/ControlNet", model_name="openpose_face")
 depth_processor = ZoeDetector.from_pretrained("lllyasviel/ControlNet")
 
+# ControlNet Modelle (bleiben im VRAM)
 controlnet_face = ControlNetModel.from_pretrained(
     "lllyasviel/control_v11p_sd15_openpose",
     subfolder="faceonly",
@@ -25,44 +21,40 @@ controlnet_depth = ControlNetModel.from_pretrained(
     torch_dtype=torch.float16
 ).to("cuda")
 
-# Pipeline-Cache (wird erst beim ersten Aufruf erstellt)
+# Pipeline-Cache
 _facefix_pipe = None
 
 def _get_facefix_pipeline(model_id: str):
     global _facefix_pipe
     if _facefix_pipe is None:
-        print(f"Lade Face-Fix-Pipeline mit Modell: {model_id}")
+        print(f"Lade Face-Fix-Pipeline für Modell: {model_id}")
         _facefix_pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
             model_id,
             controlnet=[controlnet_face, controlnet_depth],
             torch_dtype=torch.float16,
             safety_checker=None,
-            requires_safety_checker=False,
         ).to("cuda")
         _facefix_pipe.enable_xformers_memory_efficient_attention()
-        _facefix_pipe.enable_model_cpu_offload()  # spart ~2 GB bei 16 GB Karten
+        _facefix_pipe.enable_model_cpu_offload()  # spart ~2 GB!
     return _facefix_pipe
 
-def apply_facefix(image: Image.Image, prompt: str, negative_prompt: str, seed: int, base_model_path: str):
-    """
-    Automatischer 20-Sekunden-Fix für perfekte Gesichter & Hände
-    """
-    pipe = _get_facefix_pipeline(base_model_path)
+def apply_facefix(image: Image.Image, prompt: str, negative_prompt: str, seed: int, model_id: str):
+    pipe = _get_facefix_pipeline(model_id)
 
-    # Control-Images aus dem generierten Bild erzeugen
+    # Control-Images erzeugen
     pose_img = openpose_face(image)
     depth_img = depth_processor(image)
 
-    fixed_image = pipe(
+    fixed = pipe(
         prompt=prompt,
         negative_prompt=negative_prompt,
         image=image,
         control_image=[pose_img, depth_img],
-        controlnet_conditioning_scale=[0.85, 0.60],   # Face stark, Depth mittel
+        controlnet_conditioning_scale=[0.85, 0.60],
         strength=0.42,
         num_inference_steps=20,
         guidance_scale=7.0,
         generator=torch.Generator("cuda").manual_seed(seed),
     ).images[0]
 
-    return fixed_image
+    return fixed