[tests] Fix controlnet tests (#13736)

* add serge reviewer to enable claude for inline reviews.

* remove local settings

* fix controlnet test failures.

* remove serge reviewer workflow from the mix

Author: sayakpaul

tests/pipelines/controlnet_flux/test_controlnet_flux.py

@@ -143,7 +143,7 @@ def get_dummy_inputs(self, device, seed=0):
             (1, 3, 32, 32),
             generator=generator,
             device=torch.device(device),
-            dtype=torch.float16,
+            dtype=torch.float32,
         )
 
         controlnet_conditioning_scale = 0.5
@@ -163,7 +163,7 @@ def get_dummy_inputs(self, device, seed=0):
     def test_controlnet_flux(self):
         components = self.get_dummy_components()
         flux_pipe = FluxControlNetPipeline(**components)
-        flux_pipe = flux_pipe.to(torch_device, dtype=torch.float16)
+        flux_pipe = flux_pipe.to(torch_device, dtype=torch.float32)
         flux_pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(torch_device)
@@ -174,9 +174,7 @@ def test_controlnet_flux(self):
 
         assert image.shape == (1, 32, 32, 3)
 
-        expected_slice = np.array(
-            [0.47387695, 0.63134766, 0.5605469, 0.61621094, 0.7207031, 0.7089844, 0.70410156, 0.6113281, 0.64160156]
-        )
+        expected_slice = np.array([0.6677, 0.6138, 0.5296, 0.6109, 0.5672, 0.6373, 0.5463, 0.6068, 0.5569])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
             f"Expected: {expected_slice}, got: {image_slice.flatten()}"

tests/pipelines/controlnet_hunyuandit/test_controlnet_hunyuandit.py

@@ -126,7 +126,7 @@ def get_dummy_inputs(self, device, seed=0):
             (1, 3, 16, 16),
             generator=generator,
             device=torch.device(device),
-            dtype=torch.float16,
+            dtype=torch.float32,
         )
 
         controlnet_conditioning_scale = 0.5
@@ -146,7 +146,7 @@ def get_dummy_inputs(self, device, seed=0):
     def test_controlnet_hunyuandit(self):
         components = self.get_dummy_components()
         pipe = HunyuanDiTControlNetPipeline(**components)
-        pipe = pipe.to(torch_device, dtype=torch.float16)
+        pipe = pipe.to(torch_device, dtype=torch.float32)
         pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(torch_device)
@@ -156,14 +156,7 @@ def test_controlnet_hunyuandit(self):
         image_slice = image[0, -3:, -3:, -1]
         assert image.shape == (1, 16, 16, 3)
 
-        if torch_device == "xpu":
-            expected_slice = np.array(
-                [0.6948242, 0.89160156, 0.59375, 0.5078125, 0.57910156, 0.6035156, 0.58447266, 0.53564453, 0.52246094]
-            )
-        else:
-            expected_slice = np.array(
-                [0.6953125, 0.89208984, 0.59375, 0.5078125, 0.5786133, 0.6035156, 0.5839844, 0.53564453, 0.52246094]
-            )
+        expected_slice = np.array([0.5925, 0.5392, 0.4450, 0.7140, 0.3954, 0.3553, 0.3842, 0.5994, 0.3765])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
             f"Expected: {expected_slice}, got: {image_slice.flatten()}"

tests/pipelines/controlnet_sd3/test_controlnet_inpaint_sd3.py

@@ -156,14 +156,14 @@ def get_dummy_inputs(self, device, seed=0):
             (1, 3, 32, 32),
             generator=generator,
             device=torch.device(device),
-            dtype=torch.float16,
+            dtype=torch.float32,
         )
 
         control_mask = randn_tensor(
             (1, 1, 32, 32),
             generator=generator,
             device=torch.device(device),
-            dtype=torch.float16,
+            dtype=torch.float32,
         )
 
         controlnet_conditioning_scale = 0.95
@@ -184,7 +184,7 @@ def get_dummy_inputs(self, device, seed=0):
     def test_controlnet_inpaint_sd3(self):
         components = self.get_dummy_components()
         sd_pipe = StableDiffusion3ControlNetInpaintingPipeline(**components)
-        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float16)
+        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float32)
         sd_pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(torch_device)
@@ -195,9 +195,7 @@ def test_controlnet_inpaint_sd3(self):
 
         assert image.shape == (1, 32, 32, 3)
 
-        expected_slice = np.array(
-            [0.51708984, 0.7421875, 0.4580078, 0.6435547, 0.65625, 0.43603516, 0.5151367, 0.65722656, 0.60839844]
-        )
+        expected_slice = np.array([0.2875, 0.3173, 0.4028, 0.7248, 0.6338, 0.4238, 0.1730, 0.4609, 0.5424])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
             f"Expected: {expected_slice}, got: {image_slice.flatten()}"

tests/pipelines/controlnet_sd3/test_controlnet_sd3.py

@@ -173,7 +173,7 @@ def get_dummy_inputs(self, device, seed=0):
             (1, 3, 32, 32),
             generator=generator,
             device=torch.device(device),
-            dtype=torch.float16,
+            dtype=torch.float32,
         )
 
         controlnet_conditioning_scale = 0.5
@@ -192,7 +192,7 @@ def get_dummy_inputs(self, device, seed=0):
 
     def run_pipe(self, components, use_sd35=False):
         sd_pipe = StableDiffusion3ControlNetPipeline(**components)
-        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float16)
+        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float32)
         sd_pipe.set_progress_bar_config(disable=None)
 
         inputs = self.get_dummy_inputs(torch_device)
@@ -204,9 +204,9 @@ def run_pipe(self, components, use_sd35=False):
         assert image.shape == (1, 32, 32, 3)
 
         if not use_sd35:
-            expected_slice = np.array([0.5767, 0.7100, 0.5981, 0.5674, 0.5952, 0.4102, 0.5093, 0.5044, 0.6030])
+            expected_slice = np.array([0.4578, 0.3582, 0.4046, 0.0953, 0.6878, 0.5821, 0.5541, 0.5888, 0.4651])
         else:
-            expected_slice = np.array([1.0000, 0.9072, 0.4209, 0.2744, 0.5737, 0.3840, 0.6113, 0.6250, 0.6328])
+            expected_slice = np.array([0.3721, 0.5626, 0.4657, 0.2845, 0.5241, 0.5917, 0.6265, 0.6955, 0.3969])
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2, (
             f"Expected: {expected_slice}, got: {image_slice.flatten()}"