SD3 ControlNet Inpainting: MultiControlNet support

src/diffusers/pipelines/controlnet_sd3/pipeline_stable_diffusion_3_controlnet_inpainting.py

@@ -1119,7 +1119,53 @@ def __call__(
             width = latent_width * self.vae_scale_factor
 
         elif isinstance(self.controlnet, SD3MultiControlNetModel):
-            raise NotImplementedError("MultiControlNetModel is not supported for SD3ControlNetInpaintingPipeline.")
+            # Normalize inputs to lists matching number of control nets
+            num_cn = len(self.controlnet.nets)
+
+            if not isinstance(control_image, (list, tuple)):
+                control_images = [control_image] * num_cn
+            else:
+                control_images = list(control_image)
+
+            if not isinstance(control_mask, (list, tuple)):
+                control_masks = [control_mask] * num_cn
+            else:
+                control_masks = list(control_mask)
+
+            if len(control_images) != num_cn:
+                raise ValueError(
+                    f"Expected {num_cn} control images for SD3MultiControlNetModel, got {len(control_images)}."
+                )
+            if len(control_masks) != num_cn:
+                raise ValueError(
+                    f"Expected {num_cn} control masks for SD3MultiControlNetModel, got {len(control_masks)}."
+                )
+
+            # Prepare per-control inpainting conditions
+            prepared_controls = []
+            first_latent_size = None
+            for img_i, msk_i in zip(control_images, control_masks):
+                ctrl = self.prepare_image_with_mask(
+                    image=img_i,
+                    mask=msk_i,
+                    width=width,
+                    height=height,
+                    batch_size=batch_size * num_images_per_prompt,
+                    num_images_per_prompt=num_images_per_prompt,
+                    device=device,
+                    dtype=dtype,
+                    do_classifier_free_guidance=self.do_classifier_free_guidance,
+                    guess_mode=False,
+                )
+                if first_latent_size is None:
+                    first_latent_size = ctrl.shape[-2:]
+                prepared_controls.append(ctrl)
+
+            latent_height, latent_width = first_latent_size
+            height = latent_height * self.vae_scale_factor
+            width = latent_width * self.vae_scale_factor
+
+            control_image = prepared_controls
         else:
             assert False
 
@@ -1128,6 +1174,10 @@ def __call__(
         else:
             controlnet_pooled_projections = controlnet_pooled_projections or pooled_prompt_embeds
 
+        # Ensure conditioning scale broadcast for multi-control
+        if isinstance(self.controlnet, SD3MultiControlNetModel) and not isinstance(controlnet_conditioning_scale, list):
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(self.controlnet.nets)
+
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas)
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

tests/pipelines/controlnet_sd3/test_controlnet_inpaint_sd3.py

@@ -25,7 +25,7 @@
     SD3Transformer2DModel,
     StableDiffusion3ControlNetInpaintingPipeline,
 )
-from diffusers.models import SD3ControlNetModel
+from diffusers.models import SD3ControlNetModel, SD3MultiControlNetModel
 from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import (
@@ -201,3 +201,26 @@ def test_controlnet_inpaint_sd3(self):
     @unittest.skip("xFormersAttnProcessor does not work with SD3 Joint Attention")
     def test_xformers_attention_forwardGenerator_pass(self):
         pass
+
+    def test_controlnet_inpaint_sd3_multi_control(self):
+        components = self.get_dummy_components()
+        # Duplicate the single controlnet into a MultiControlNet for smoke test
+        cn = components["controlnet"]
+        components["controlnet"] = SD3MultiControlNetModel([cn, cn])
+
+        sd_pipe = StableDiffusion3ControlNetInpaintingPipeline(**components)
+        sd_pipe = sd_pipe.to(torch_device, dtype=torch.float16)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        # Provide lists for multi-control
+        inputs["control_image"] = [inputs["control_image"], inputs["control_image"]]
+        inputs["control_mask"] = [inputs["control_mask"], inputs["control_mask"]]
+        inputs["controlnet_conditioning_scale"] = [1.0, 0.5]
+        inputs["num_inference_steps"] = 2
+
+        output = sd_pipe(**inputs)
+        image = output.images
+
+        # Shape check only (deterministic slice check not required for multi-control smoke test)
+        assert image.shape == (1, 32, 32, 3)