diff --git a/src/transformers/models/hunyuan_vl/modeling_hunyuan_vl.py b/src/transformers/models/hunyuan_vl/modeling_hunyuan_vl.py
index 00d78217df68..26d945b647ed 100644
--- a/src/transformers/models/hunyuan_vl/modeling_hunyuan_vl.py
+++ b/src/transformers/models/hunyuan_vl/modeling_hunyuan_vl.py
@@ -979,7 +979,12 @@ def get_image_features(
     ) -> tuple | BaseModelOutputWithPooling:
         vision_dtype = next(self.vision_tower.parameters()).dtype
         pixel_values = pixel_values.to(vision_dtype)
-        return self.vision_tower(pixel_values, grid_thw=image_grid_thw, **kwargs)
+        image_outputs = self.vision_tower(pixel_values, grid_thw=image_grid_thw, **kwargs)
+        # Include row separators and image boundary tokens in each split.
+        merged_h, merged_w = (image_grid_thw[:, 1:] // self.config.vision_config.spatial_merge_size).unbind(-1)
+        split_sizes = (merged_h * (merged_w + 1) + 2).tolist()
+        image_outputs.pooler_output = torch.split(image_outputs.pooler_output[0], split_sizes)
+        return image_outputs
 
     def get_placeholder_mask(
         self,
@@ -1064,7 +1069,9 @@ def forward(
             mm_encoder_outputs["image"] = self.get_image_features(pixel_values, image_grid_thw, return_dict=True)
 
         if mm_encoder_outputs.get("image") is not None:
-            image_embeds = mm_encoder_outputs["image"].pooler_output.to(inputs_embeds.device, inputs_embeds.dtype)
+            image_embeds = torch.cat(mm_encoder_outputs["image"].pooler_output, dim=0).to(
+                inputs_embeds.device, inputs_embeds.dtype
+            )
             image_mask = self.get_placeholder_mask(
                 input_ids,
                 inputs_embeds=inputs_embeds,
diff --git a/src/transformers/models/hunyuan_vl/modular_hunyuan_vl.py b/src/transformers/models/hunyuan_vl/modular_hunyuan_vl.py
index 047fba93c4ca..b33c0f6d3f9c 100644
--- a/src/transformers/models/hunyuan_vl/modular_hunyuan_vl.py
+++ b/src/transformers/models/hunyuan_vl/modular_hunyuan_vl.py
@@ -1224,7 +1224,12 @@ def get_image_features(
     ) -> tuple | BaseModelOutputWithPooling:
         vision_dtype = next(self.vision_tower.parameters()).dtype
         pixel_values = pixel_values.to(vision_dtype)
-        return self.vision_tower(pixel_values, grid_thw=image_grid_thw, **kwargs)
+        image_outputs = self.vision_tower(pixel_values, grid_thw=image_grid_thw, **kwargs)
+        # Include row separators and image boundary tokens in each split.
+        merged_h, merged_w = (image_grid_thw[:, 1:] // self.config.vision_config.spatial_merge_size).unbind(-1)
+        split_sizes = (merged_h * (merged_w + 1) + 2).tolist()
+        image_outputs.pooler_output = torch.split(image_outputs.pooler_output[0], split_sizes)
+        return image_outputs
 
     def get_placeholder_mask(
         self,
@@ -1281,7 +1286,9 @@ def forward(
             mm_encoder_outputs["image"] = self.get_image_features(pixel_values, image_grid_thw, return_dict=True)
 
         if mm_encoder_outputs.get("image") is not None:
-            image_embeds = mm_encoder_outputs["image"].pooler_output.to(inputs_embeds.device, inputs_embeds.dtype)
+            image_embeds = torch.cat(mm_encoder_outputs["image"].pooler_output, dim=0).to(
+                inputs_embeds.device, inputs_embeds.dtype
+            )
             image_mask = self.get_placeholder_mask(
                 input_ids,
                 inputs_embeds=inputs_embeds,
diff --git a/src/transformers/models/idefics2/modeling_idefics2.py b/src/transformers/models/idefics2/modeling_idefics2.py
index c73814f645fd..d1ff9f92d848 100644
--- a/src/transformers/models/idefics2/modeling_idefics2.py
+++ b/src/transformers/models/idefics2/modeling_idefics2.py
@@ -836,6 +836,7 @@ def get_image_features(
         # Remove padding images - padding images are full 0.
         nb_values_per_image = pixel_values.shape[1:].numel()
         real_images_inds = (pixel_values == 0.0).sum(dim=(-1, -2, -3)) != nb_values_per_image
+        num_real_images_per_batch = real_images_inds.view(batch_size, num_images).sum(dim=1)
         pixel_values = pixel_values[real_images_inds].contiguous()
 
         # Handle the vision attention mask
@@ -861,9 +862,12 @@ def get_image_features(
         image_hidden_states = image_outputs.last_hidden_state
 
         # Modality projection & resampling
-        image_outputs.pooler_output = self.connector(
+        image_features = self.connector(
             image_hidden_states, attention_mask=patch_attention_mask.view(pixel_values.size(0), -1)
         )
+        image_outputs.pooler_output = [
+            f.reshape(-1, f.shape[-1]) for f in image_features.split(num_real_images_per_batch.tolist())
+        ]
 
         return image_outputs
 
@@ -938,6 +942,8 @@ def forward(
             image_hidden_states = None
 
         if image_hidden_states is not None:
+            if isinstance(image_hidden_states, (list, tuple)):
+                image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=inputs_embeds.device)
             # When we generate, we don't want to replace the potential image_token_id that we generated by images
             # that simply don't exist
diff --git a/src/transformers/models/idefics3/modeling_idefics3.py b/src/transformers/models/idefics3/modeling_idefics3.py
index 3ef9f8917e16..e66243911d3d 100644
--- a/src/transformers/models/idefics3/modeling_idefics3.py
+++ b/src/transformers/models/idefics3/modeling_idefics3.py
@@ -580,6 +580,7 @@ def get_image_features(
         # Remove padding images - padding images are full 0.
         nb_values_per_image = pixel_values.shape[1:].numel()
         real_images_inds = (pixel_values == 0.0).sum(dim=(-1, -2, -3)) != nb_values_per_image
+        num_real_images_per_batch = real_images_inds.view(batch_size, num_images).sum(dim=1)
         pixel_values = pixel_values[real_images_inds].contiguous()
 
         # Handle the vision attention mask
@@ -607,7 +608,9 @@ def get_image_features(
 
         # Modality projection & resampling
         image_features = self.connector(image_hidden_states)
-        image_outputs.pooler_output = image_features
+        image_outputs.pooler_output = [
+            f.reshape(-1, f.shape[-1]) for f in image_features.split(num_real_images_per_batch.tolist())
+        ]
 
         return image_outputs
 
@@ -679,6 +682,8 @@ def forward(
             image_hidden_states = None
 
         if image_hidden_states is not None:
+            if isinstance(image_hidden_states, (list, tuple)):
+                image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=inputs_embeds.device)
             # When we generate, we don't want to replace the potential image_token_id that we generated by images
             # that simply don't exist
diff --git a/src/transformers/models/llava_next_video/modeling_llava_next_video.py b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
index 18c42b938bb7..dbe8fab9800e 100644
--- a/src/transformers/models/llava_next_video/modeling_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
@@ -640,7 +640,7 @@ def get_video_features(
         # Same as image features except that video has pooling layer
         video_features = self.vision_resampler(selected_video_features)
         video_features = self.multi_modal_projector(video_features)
-        video_outputs.pooler_output = torch.split(video_features, frames, dim=0)
+        video_outputs.pooler_output = video_features.unflatten(0, (batch_size, frames)).flatten(1, 2)
         return video_outputs
 
 
diff --git a/src/transformers/models/llava_next_video/modular_llava_next_video.py b/src/transformers/models/llava_next_video/modular_llava_next_video.py
index afddc98dea4b..b272b3ca990c 100644
--- a/src/transformers/models/llava_next_video/modular_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/modular_llava_next_video.py
@@ -339,7 +339,7 @@ def get_video_features(
         # Same as image features except that video has pooling layer
         video_features = self.vision_resampler(selected_video_features)
         video_features = self.multi_modal_projector(video_features)
-        video_outputs.pooler_output = torch.split(video_features, frames, dim=0)
+        video_outputs.pooler_output = video_features.unflatten(0, (batch_size, frames)).flatten(1, 2)
         return video_outputs
 
     def get_placeholder_mask(
diff --git a/src/transformers/models/modernvbert/modeling_modernvbert.py b/src/transformers/models/modernvbert/modeling_modernvbert.py
index ef329f6b7475..0f8bd63eccc8 100755
--- a/src/transformers/models/modernvbert/modeling_modernvbert.py
+++ b/src/transformers/models/modernvbert/modeling_modernvbert.py
@@ -241,7 +241,8 @@ def inputs_merger(
         - The merging happens so that we obtain the following sequence: `vector_tok_1 vector_tok_2 vector_tok_3 vector_fake_tok_around_image {sequence of image_seq_len image hidden states} vector_fake_toke_around_image vector_tok_4`. That sequence is fed to the LM.
         - To fit the format of that sequence, `input_ids`, `inputs_embeds`, `attention_mask` are all 3 adapted to insert the image hidden states.
         """
-        _, patch_size, _ = image_hidden_states.shape
+        patch_size = self.config.vision_config.patch_size
+        image_hidden_states = image_hidden_states.view(-1, patch_size, image_hidden_states.shape[-1])
 
         if input_ids is None:
             image_mask = inputs_embeds == self.get_input_embeddings()(
@@ -297,6 +298,7 @@ def get_image_features(
 
         # If no images, leave one empty image.
         real_images_inds[0] |= ~torch.any(real_images_inds)
+        num_real_images_per_batch = real_images_inds.view(batch_size, num_images).sum(dim=1)
 
         pixel_values = pixel_values[real_images_inds].contiguous()
         # Handle the vision attention mask
@@ -323,7 +325,9 @@ def get_image_features(
 
         # Modality projection & resampling
         image_features = self.connector(image_hidden_states)
-        image_outputs.pooler_output = image_features
+        image_outputs.pooler_output = [
+            f.reshape(-1, f.shape[-1]) for f in image_features.split(num_real_images_per_batch.tolist())
+        ]
 
         return image_outputs
 
@@ -369,6 +373,7 @@ def forward(
 
         # Merge image and text embeddings
         if image_hidden_states is not None:
+            image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
             inputs_embeds = self.inputs_merger(
                 input_ids=input_ids, inputs_embeds=inputs_embeds, image_hidden_states=image_hidden_states
diff --git a/src/transformers/models/modernvbert/modular_modernvbert.py b/src/transformers/models/modernvbert/modular_modernvbert.py
index ed1a8966d492..65329c1fe7f5 100755
--- a/src/transformers/models/modernvbert/modular_modernvbert.py
+++ b/src/transformers/models/modernvbert/modular_modernvbert.py
@@ -296,6 +296,7 @@ def forward(
 
         # Merge image and text embeddings
         if image_hidden_states is not None:
+            image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
             inputs_embeds = self.inputs_merger(
                 input_ids=input_ids, inputs_embeds=inputs_embeds, image_hidden_states=image_hidden_states
diff --git a/src/transformers/models/smolvlm/modeling_smolvlm.py b/src/transformers/models/smolvlm/modeling_smolvlm.py
index 2172e97610bc..d30707851096 100644
--- a/src/transformers/models/smolvlm/modeling_smolvlm.py
+++ b/src/transformers/models/smolvlm/modeling_smolvlm.py
@@ -478,7 +478,8 @@ def inputs_merger(
         - The merging happens so that we obtain the following sequence: `vector_tok_1 vector_tok_2 vector_tok_3 vector_fake_tok_around_image {sequence of image_seq_len image hidden states} vector_fake_toke_around_image vector_tok_4`. That sequence is fed to the LM.
         - To fit the format of that sequence, `input_ids`, `inputs_embeds`, `attention_mask` are all 3 adapted to insert the image hidden states.
         """
-        _, patch_size, _ = image_hidden_states.shape
+        patch_size = self.config.vision_config.patch_size
+        image_hidden_states = image_hidden_states.view(-1, patch_size, image_hidden_states.shape[-1])
 
         if input_ids is None:
             image_mask = inputs_embeds == self.get_input_embeddings()(
@@ -534,6 +535,7 @@ def get_image_features(
 
         # If no images, leave one empty image.
         real_images_inds[0] |= ~torch.any(real_images_inds)
+        num_real_images_per_batch = real_images_inds.view(batch_size, num_images).sum(dim=1)
 
         pixel_values = pixel_values[real_images_inds].contiguous()
         # Handle the vision attention mask
@@ -560,7 +562,9 @@ def get_image_features(
 
         # Modality projection & resampling
         image_features = self.connector(image_hidden_states)
-        image_outputs.pooler_output = image_features
+        image_outputs.pooler_output = [
+            f.reshape(-1, f.shape[-1]) for f in image_features.split(num_real_images_per_batch.tolist())
+        ]
 
         return image_outputs
 
@@ -624,6 +628,8 @@ def forward(
             image_hidden_states = None
 
         if image_hidden_states is not None:
+            if isinstance(image_hidden_states, (list, tuple)):
+                image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=inputs_embeds.device)
             # When we generate, we don't want to replace the potential image_token_id that we generated by images
             # that simply don't exist
diff --git a/src/transformers/models/smolvlm/modular_smolvlm.py b/src/transformers/models/smolvlm/modular_smolvlm.py
index 17c19fdf1c2a..b6377cfd129b 100644
--- a/src/transformers/models/smolvlm/modular_smolvlm.py
+++ b/src/transformers/models/smolvlm/modular_smolvlm.py
@@ -110,7 +110,8 @@ class SmolVLMModel(Idefics3Model):
     def inputs_merger(
         self, input_ids: torch.LongTensor, inputs_embeds: torch.Tensor, image_hidden_states: torch.Tensor
     ):
-        _, patch_size, _ = image_hidden_states.shape
+        patch_size = self.config.vision_config.patch_size
+        image_hidden_states = image_hidden_states.view(-1, patch_size, image_hidden_states.shape[-1])
 
         if input_ids is None:
             image_mask = inputs_embeds == self.get_input_embeddings()(
@@ -166,6 +167,7 @@ def get_image_features(
 
         # If no images, leave one empty image.
         real_images_inds[0] |= ~torch.any(real_images_inds)
+        num_real_images_per_batch = real_images_inds.view(batch_size, num_images).sum(dim=1)
 
         pixel_values = pixel_values[real_images_inds].contiguous()
         # Handle the vision attention mask
@@ -192,7 +194,9 @@ def get_image_features(
 
         # Modality projection & resampling
         image_features = self.connector(image_hidden_states)
-        image_outputs.pooler_output = image_features
+        image_outputs.pooler_output = [
+            f.reshape(-1, f.shape[-1]) for f in image_features.split(num_real_images_per_batch.tolist())
+        ]
 
         return image_outputs
 
@@ -252,6 +256,8 @@ def forward(
             image_hidden_states = None
 
         if image_hidden_states is not None:
+            if isinstance(image_hidden_states, (list, tuple)):
+                image_hidden_states = torch.cat(image_hidden_states, dim=0)
             image_hidden_states = image_hidden_states.to(dtype=self.dtype, device=inputs_embeds.device)
             # When we generate, we don't want to replace the potential image_token_id that we generated by images
             # that simply don't exist
diff --git a/tests/models/chameleon/test_modeling_chameleon.py b/tests/models/chameleon/test_modeling_chameleon.py
index 2ca994b1569d..0e97eb674404 100644
--- a/tests/models/chameleon/test_modeling_chameleon.py
+++ b/tests/models/chameleon/test_modeling_chameleon.py
@@ -245,6 +245,10 @@ def test_batching_equivalence(self):
     def test_get_image_features_output(self, return_dict: bool | None):
         pass
 
+    @unittest.skip("Skip get_image_features tests as those are tested via ChameleonVision2SeqModelTest instead")
+    def test_get_image_features_pooler_output(self):
+        pass
+
     @unittest.skip("Skip get_image_features tests as those are tested via ChameleonVision2SeqModelTest instead")
     def test_get_image_features_hidden_states(self):
         pass
diff --git a/tests/models/gemma4/test_modeling_gemma4.py b/tests/models/gemma4/test_modeling_gemma4.py
index f0e1e812e451..80504e56c399 100644
--- a/tests/models/gemma4/test_modeling_gemma4.py
+++ b/tests/models/gemma4/test_modeling_gemma4.py
@@ -288,6 +288,10 @@ def test_get_image_features_attentions(self):
     def test_get_image_features_output(self, return_dict: bool | None):
         pass
 
+    @unittest.skip("The tester has no image in input dict")
+    def test_get_image_features_pooler_output(self):
+        pass
+
     @unittest.skip("The tester has no videos in input dict")
     def test_get_video_features_hidden_states(self):
         pass
diff --git a/tests/models/gemma4_unified/test_modeling_gemma4_unified.py b/tests/models/gemma4_unified/test_modeling_gemma4_unified.py
index 78fdde7b0720..08bb08c7c624 100644
--- a/tests/models/gemma4_unified/test_modeling_gemma4_unified.py
+++ b/tests/models/gemma4_unified/test_modeling_gemma4_unified.py
@@ -244,6 +244,7 @@ def skip_mm_output_format(self):
             "test_get_image_features_hidden_states",
             "test_get_image_features_attentions",
             "test_get_image_features_output",
+            "test_get_image_features_pooler_output",
             "test_get_video_features_hidden_states",
             "test_get_video_features_attentions",
             "test_get_video_features_output",
diff --git a/tests/models/inkling/test_modeling_inkling.py b/tests/models/inkling/test_modeling_inkling.py
index a7b28f8cd59f..fa08aff6488b 100644
--- a/tests/models/inkling/test_modeling_inkling.py
+++ b/tests/models/inkling/test_modeling_inkling.py
@@ -176,6 +176,10 @@ def test_get_image_features_attentions(self):
     def test_get_image_features_output(self, return_dict: bool | None):
         pass
 
+    @unittest.skip("The tester has no image in input dict")
+    def test_get_image_features_pooler_output(self):
+        pass
+
     @unittest.skip("The tester has no videos in input dict")
     def test_get_video_features_hidden_states(self):
         pass
diff --git a/tests/models/nemotron_h_omni/test_modeling_nemotron_h_omni.py b/tests/models/nemotron_h_omni/test_modeling_nemotron_h_omni.py
index fbce3a88e74c..98b9294bcc13 100644
--- a/tests/models/nemotron_h_omni/test_modeling_nemotron_h_omni.py
+++ b/tests/models/nemotron_h_omni/test_modeling_nemotron_h_omni.py
@@ -347,6 +347,11 @@ def test_get_image_features_output(self, return_dict: bool | None):
     def test_get_video_features_output(self, return_dict: bool | None):
         pass
 
+    @unittest.skip("FIXME raushan - should apply temporal factor while processing!")
+    # prob this model doesn't even work with beam/sampling atm, low prio
+    def test_get_video_features_pooler_output(self):
+        pass
+
     def test_mismatching_num_image_tokens(self):
         # packed `pixel_values` have no batch dimension, so images are dropped or added by their patches
         config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -435,6 +440,11 @@ def test_get_image_features_output(self, return_dict: bool | None):
     def test_get_video_features_output(self, return_dict: bool | None):
         pass
 
+    @unittest.skip("FIXME raushan - should apply temporal factor while processing!")
+    # prob this model doesn't even work with beam/sampling atm, low prio
+    def test_get_video_features_pooler_output(self):
+        pass
+
     @unittest.skip(reason="Audio tester isnt supported yet")
     def test_generate_from_multimodal_encoder_outputs_and_raw_data(self):
         pass
diff --git a/tests/test_modeling_common.py b/tests/test_modeling_common.py
index 80dcbd30bd60..5c65eade0373 100644
--- a/tests/test_modeling_common.py
+++ b/tests/test_modeling_common.py
@@ -5407,24 +5407,38 @@ def test_get_image_features_output(self, return_dict: bool | None):
                     hidden_size,
                     f"hidden_size mismatch, full shape: {last_hidden_state_shape}",
                 )
-
-                if (
-                    model_class.__name__ in get_values(MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES)
-                    and outputs.pooler_output is not None
-                ):
-                    self.assertEqual(
-                        len(outputs.pooler_output),
-                        self.model_tester.batch_size,
-                        f"batch_size mismatch for `pooler_output`: {len(outputs.pooler_output)} != {self.model_tester.batch_size}",
-                    )
-                    self.assertEqual(
-                        outputs.pooler_output[0].ndim,
-                        2,
-                        f"each sample in `pooler_output` should be a 2D array but got {outputs.pooler_output[0].ndim}",
-                    )
             else:
                 self.assertIsInstance(outputs, tuple, "get_image_features() must return a tuple if return_dict=False")
 
+    def test_get_image_features_pooler_output(self):
+        """
+        Generation assumes that encoded images always follow certain layout. This test is IMPORTANT
+        and should NEVER be skipped unless your model override beam-search expansion on `generate()`
+        """
+        for model_class in self.all_model_classes:
+            if model_class.__name__ not in get_values(MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES):
+                continue
+            if not hasattr(model_class, "get_image_features"):
+                continue
+
+            config, inputs_dict = self._image_features_prepare_config_and_inputs()
+            model = model_class(config).eval().to(torch_device)
+            with torch.no_grad():
+                pooler_output = model.get_image_features(**inputs_dict, return_dict=True).pooler_output
+            if pooler_output is None:
+                continue
+
+            self.assertEqual(
+                len(pooler_output),
+                self.model_tester.batch_size,
+                f"`pooler_output` should have one entry per image: {len(pooler_output)} != {self.model_tester.batch_size}",
+            )
+            self.assertEqual(
+                pooler_output[0].ndim,
+                2,
+                f"each sample in `pooler_output` should be a 2D array but got {pooler_output[0].ndim}",
+            )
+
     def test_get_image_features_hidden_states(self):
         def check_hidden_states_output(inputs_dict, config, model_class):
             model = model_class(copy.deepcopy(config))
@@ -5724,6 +5738,35 @@ def test_get_video_features_output(self, return_dict: bool | None):
             else:
                 self.assertIsInstance(outputs, tuple, "get_video_features() must return a tuple if return_dict=False")
 
+    def test_get_video_features_pooler_output(self):
+        """
+        Generation assumes that encoded images always follow certain layout. This test is IMPORTANT
+        and should NEVER be skipped unless your model override beam-search expansion on `generate()`
+        """
+        for model_class in self.all_model_classes:
+            if model_class.__name__ not in get_values(MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES):
+                continue
+            if not hasattr(model_class, "get_video_features"):
+                continue
+
+            config, inputs_dict = self._video_features_prepare_config_and_inputs()
+            model = model_class(config).eval().to(torch_device)
+            with torch.no_grad():
+                pooler_output = model.get_video_features(**inputs_dict, return_dict=True).pooler_output
+            if pooler_output is None:
+                continue
+
+            self.assertEqual(
+                len(pooler_output),
+                self.model_tester.batch_size,
+                f"batch_size mismatch for `pooler_output`: {len(pooler_output)} != {self.model_tester.batch_size}",
+            )
+            self.assertEqual(
+                pooler_output[0].ndim,
+                2,
+                f"each sample in `pooler_output` should be a 2D array but got {pooler_output[0].ndim}",
+            )
+
     def test_get_video_features_hidden_states(self):
         def check_hidden_states_output(inputs_dict, config, model_class):
             model = model_class(copy.deepcopy(config))