fix(cpu): replace CUDA-only calls; use DDIM for remove_hair on CPU/GPU

Hair_stable_new_fresh/infer_full.py

@@ -119,7 +119,11 @@ class StableHair:
             safety_checker=None,
             torch_dtype=weight_dtype,
         )
-        self.remove_hair_pipeline.scheduler = UniPCMultistepScheduler.from_config(self.remove_hair_pipeline.scheduler.config)
+        # UniPC can throw 'stack expects a non-empty TensorList' or order-related
+        # assertions on some CPU builds. Use DDIM for wider compatibility.
+        self.remove_hair_pipeline.scheduler = DDIMScheduler.from_config(
+            self.remove_hair_pipeline.scheduler.config
+        )
         self.remove_hair_pipeline = self.remove_hair_pipeline.to(device)
 
         ### move to fp16

utils/pipeline.py

@@ -275,7 +275,7 @@ class StableHairPipeline(DiffusionPipeline):
             condition = condition
         elif isinstance(condition, np.ndarray):
             # suppose input is [0, 255]
-            condition = self.images2latents(condition, dtype).cuda()
+            condition = self.images2latents(condition, dtype).to(device)
         if do_classifier_free_guidance:
             condition_pad = torch.ones_like(condition) * -1
             condition = torch.cat([condition_pad, condition])
@@ -423,12 +423,13 @@ class StableHairPipeline(DiffusionPipeline):
             num_actual_inference_steps = num_inference_steps
 
         if isinstance(ref_image, str):
-            ref_image_latents = self.images2latents(np.array(Image.open(ref_image).resize((width, height))),
-                                                    latents_dtype).cuda()
+            ref_image_latents = self.images2latents(
+                np.array(Image.open(ref_image).resize((width, height))), latents_dtype
+            ).to(device)
         elif isinstance(ref_image, np.ndarray):
-            ref_image_latents = self.images2latents(ref_image, latents_dtype).cuda()
+            ref_image_latents = self.images2latents(ref_image, latents_dtype).to(device)
         elif isinstance(ref_image, torch.Tensor):
-            ref_image_latents = self.images2latents(ref_image, latents_dtype).cuda()
+            ref_image_latents = self.images2latents(ref_image, latents_dtype).to(device)
 
         ref_padding_latents = torch.ones_like(ref_image_latents) * -1
         ref_image_latents = torch.cat([ref_padding_latents, ref_image_latents]) if do_classifier_free_guidance else ref_image_latents