Fix: use s_scale=0 when IP-Adapter loaded but no face requested

When IP-Adapter FaceID is initialized, it modifies the pipeline's UNet
cross-attention layers. Calling raw pipeline() without face embeddings
leaves these layers in a broken state, causing corrupted output.

Solution: When IP-Adapter is loaded but no face_image provided, call
ip_model.generate() with s_scale=0.0 and zero embeddings to properly
disable face conditioning while satisfying the modified layers.

requirements.txt

@@ -28,6 +28,7 @@ insightface==0.7.3
 
 # IP-Adapter from GitHub
 git+https://github.com/tencent-ailab/IP-Adapter.git
+einops  # Required by IP-Adapter
 
 # Utilities
 pydantic==2.6.0

worker/generator.py

@@ -289,7 +289,30 @@ class SDXLGenerator:
                     )[0]
                 )
             else:
-                # Progress callback wrapper (only for standard pipeline)
+                # Check if IP-Adapter is loaded - if so, we must use it with s_scale=0
+                # to avoid corrupted output from dangling adapter layers
+                if self.ip_adapter_loaded:
+                    logger.info("IP-Adapter loaded but no face requested, using s_scale=0")
+                    # Create zero embedding (512-dim for FaceID)
+                    zero_embed = torch.zeros((1, 512), device=self.device, dtype=torch.float16)
+                    
+                    image = await loop.run_in_executor(
+                        None,
+                        lambda: self.ip_model.generate(
+                            prompt=prompt,
+                            negative_prompt=negative_prompt,
+                            faceid_embeds=zero_embed,
+                            width=width,
+                            height=height,
+                            num_inference_steps=num_inference_steps,
+                            guidance_scale=guidance_scale,
+                            num_samples=1,
+                            seed=seed,
+                            s_scale=0.0,  # Disable face conditioning
+                        )[0]
+                    )
+                else:
+                    # Standard generation - IP-Adapter not loaded
                     def callback_wrapper(step: int, timestep: int, latents: torch.FloatTensor):
                         if progress_callback:
                             progress = int((step / num_inference_steps) * 100)
@@ -298,7 +321,6 @@ class SDXLGenerator:
                             except:
                                 pass
 
-                # Standard generation without face lock
                     image = await loop.run_in_executor(
                         None,
                         lambda: self.pipeline(