Update models.py

models.py

@@ -1,7 +1,7 @@
 """
 Models.py - Following examplewithface.py EXACTLY
 NO MultiControlNetModel wrapper!
-NO fuse_lora with scale!
+Using fuse_lora with scale (examplewithface.py line 267)
 """
 import torch
 import time
@@ -119,7 +119,7 @@ def load_sdxl_pipeline(controlnets):
     pipe = StableDiffusionXLInstantIDImg2ImgPipeline.from_pretrained(
         "frankjoshua/albedobaseXL_v21",
         vae=vae,
-        controlnet=controlnets,  # ← LIST [identitynet, zoedepthnet] - NO WRAPPER!
+        controlnet=controlnets,  # ← LIST [identitynet, zoedepthnet] - NO WRAPPER!
         torch_dtype=dtype
     )
     print("  [OK] Pipeline created with direct controlnet list")
@@ -160,7 +160,8 @@ def load_lora(pipe):
 
 def fuse_lora_with_scale(pipe, lora_scale):
     """
-    Modern approach: Load LoRA and let cross_attention_kwargs apply scale
+    Following examplewithface.py lines 266-267:
+    Load LoRA weights and FUSE them into the model
     """
     global lora_path_cached
     
@@ -168,20 +169,27 @@ def fuse_lora_with_scale(pipe, lora_scale):
         return False
     
     try:
-        # Unload previous
+        # Unload and unfuse previous LoRA if exists
         try:
+            pipe.unfuse_lora()
             pipe.unload_lora_weights()
         except:
             pass
         
-        # Load LoRA
-        print(f"  [LORA] Loading with scale {lora_scale}...")
+        # Load LoRA weights (examplewithface.py line 266)
+        print(f"  [LORA] Loading weights...")
         pipe.load_lora_weights(lora_path_cached)
-        print(f"  [OK] LoRA loaded (scale will be applied via cross_attention_kwargs)")
+        
+        # CRITICAL: Fuse LoRA into model (examplewithface.py line 267)
+        print(f"  [LORA] Fusing with scale {lora_scale}...")
+        pipe.fuse_lora(lora_scale)
+        print(f"  [OK] LoRA fused into model")
         
         return True
     except Exception as e:
-        print(f"  [ERROR] LoRA failed: {e}")
+        print(f"  [ERROR] LoRA fusion failed: {e}")
+        import traceback
+        traceback.print_exc()
         return False
 
 
@@ -363,7 +371,7 @@ def setup_ip_adapter(pipe, image_encoder):
         
         print("  [OK] IP-Adapter fully loaded with InstantID architecture")
         print(f"  - Resampler: 4 layers, 20 heads, 16 output tokens")
-        print(f"  - Face embeddings: 512D → 16x2048D")
+        print(f"  - Face embeddings: 512D → 16x2048D")
         
         return image_proj_model, True