first commit

tests/models/lfm2_vl/test_image_processing_lfm2_vl.py

@@ -0,0 +1,903 @@
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_torch_available():
+    import torch
+
+    from transformers.models.lfm2_vl.image_processing_lfm2_vl import (
+        find_closest_aspect_ratio,
+        round_by_factor,
+    )
+
+
+class Lfm2VlImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        num_images=1,
+        min_resolution=256,
+        max_resolution=1024,
+        downsample_factor=2,
+        do_image_splitting=False,
+        min_tiles=2,
+        max_tiles=10,
+        use_thumbnail=True,
+        min_image_tokens=64,
+        max_image_tokens=256,
+        encoder_patch_size=16,
+        tile_size=512,
+        max_pixels_tolerance=2.0,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.num_images = num_images
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+
+        self.downsample_factor = downsample_factor
+        self.do_image_splitting = do_image_splitting
+        self.min_tiles = min_tiles
+        self.max_tiles = max_tiles
+        self.use_thumbnail = use_thumbnail
+        self.min_image_tokens = min_image_tokens
+        self.max_image_tokens = max_image_tokens
+        self.encoder_patch_size = encoder_patch_size
+        self.tile_size = tile_size
+        self.max_pixels_tolerance = max_pixels_tolerance
+
+    def prepare_image_processor_dict(self):
+        return {
+            "downsample_factor": self.downsample_factor,
+            "do_image_splitting": self.do_image_splitting,
+            "min_tiles": self.min_tiles,
+            "max_tiles": self.max_tiles,
+            "use_thumbnail": self.use_thumbnail,
+            "min_image_tokens": self.min_image_tokens,
+            "max_image_tokens": self.max_image_tokens,
+            "encoder_patch_size": self.encoder_patch_size,
+            "tile_size": self.tile_size,
+            "max_pixels_tolerance": self.max_pixels_tolerance,
+        }
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        images = prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+        return [[image] for image in images]
+
+
+@require_torch
+@require_vision
+class Lfm2VlImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = Lfm2VlImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processing_classes.values():
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "downsample_factor"))
+            self.assertTrue(hasattr(image_processing, "min_tiles"))
+            self.assertTrue(hasattr(image_processing, "max_tiles"))
+            self.assertTrue(hasattr(image_processing, "use_thumbnail"))
+            self.assertTrue(hasattr(image_processing, "min_image_tokens"))
+            self.assertTrue(hasattr(image_processing, "max_image_tokens"))
+            self.assertTrue(hasattr(image_processing, "encoder_patch_size"))
+            self.assertTrue(hasattr(image_processing, "tile_size"))
+            self.assertTrue(hasattr(image_processing, "max_pixels_tolerance"))
+
+    @require_vision
+    def test_smart_resize(self):
+        # verify that smart resize output dims are divisible by encoder_patch_size * downsample_factor
+        image_processing = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        width, height = image_processing.smart_resize(
+            height=500,
+            width=300,
+            downsample_factor=image_processing.downsample_factor,
+            min_image_tokens=image_processing.min_image_tokens,
+            max_image_tokens=image_processing.max_image_tokens,
+            encoder_patch_size=image_processing.encoder_patch_size,
+        )
+        mod = image_processing.encoder_patch_size * image_processing.downsample_factor
+        self.assertEqual(width % mod, 0)
+        self.assertEqual(height % mod, 0)
+
+    @require_vision
+    def test_get_grid_layout(self):
+        # splitting a 512×512 image into tiles of size processor.image_processor.tile_size
+        image_processing = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        rows, cols, _, _, num_patches = image_processing._get_grid_layout(
+            height=1024,
+            width=1024,
+            min_tiles=image_processing.min_tiles,
+            max_tiles=image_processing.max_tiles,
+            tile_size=image_processing.tile_size,
+        )
+        self.assertEqual(num_patches, 4)
+        self.assertEqual(num_patches, rows * cols)
+
+        rows, cols, _, _, num_patches = image_processing._get_grid_layout(
+            height=1024,
+            width=1024,
+            min_tiles=8,
+            max_tiles=8,
+            tile_size=image_processing.tile_size,
+        )
+        self.assertEqual(num_patches, 8)
+        self.assertEqual(num_patches, rows * cols)
+
+    def test_find_closest_aspect_ratio(self):
+        # should pick (1,1) over (2,1) for a square image
+        result = find_closest_aspect_ratio(1.0, [(1, 1), (2, 1)], width=100, height=100, image_size=100)
+        self.assertEqual(result, (1, 1))
+
+        result = find_closest_aspect_ratio(0.5, [(1, 1), (1, 2)], width=100, height=200, image_size=200)
+        self.assertEqual(result, (1, 2))
+
+    def test_call_numpy(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+        for sample_images in image_inputs:
+            for image in sample_images:
+                self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_numpy_4_channels(self):
+        # Lfm2Vl always processes images as RGB, so it always returns images with 3 channels
+        # Initialize image_processing
+        image_processor_dict = self.image_processor_dict
+        image_processing = self.image_processing_classes["torchvision"](**image_processor_dict)
+        # create random numpy tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+
+        for sample_images in image_inputs:
+            for image in sample_images:
+                self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_pil(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        # create random PIL images
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize image_processing
+        image_processing = self.image_processing_classes["torchvision"](**self.image_processor_dict)
+        # create random PyTorch tensors
+        image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+        for images in image_inputs:
+            for image in images:
+                self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (1, image_processing.max_num_patches, 3 * image_processing.encoder_patch_size**2),
+        )
+
+        # Test batched
+        encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            tuple(encoded_images.shape),
+            (
+                self.image_processor_tester.batch_size,
+                image_processing.max_num_patches,
+                3 * image_processing.encoder_patch_size**2,
+            ),
+        )
+
+    def test_small_image_no_tiling_no_thumbnail(self):
+        """Small image with tiling disabled should use smart resize, no thumbnail."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=False,
+            use_thumbnail=True,  # even if enabled, should not be used for small/non-tiled images
+        )
+        # Create a small image (256x256)
+        small_image = Image.new("RGB", (256, 256), color="red")
+        result = image_processing([[small_image]], return_tensors="pt", return_row_col_info=True)
+
+        # With tiling disabled, should be 1 tile (no thumbnail)
+        self.assertEqual(result.image_rows[0].item(), 1)
+        self.assertEqual(result.image_cols[0].item(), 1)
+        # Should have exactly 1 image in batch (no thumbnail)
+        self.assertEqual(result.pixel_values.shape[0], 1)
+
+    def test_small_image_tiling_enabled_no_thumbnail(self):
+        """Small image with tiling enabled should not be tiled (too small), no thumbnail."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=10,
+        )
+        # Create a small image that won't exceed the max_image_tokens threshold
+        small_image = Image.new("RGB", (256, 256), color="blue")
+        result = image_processing([[small_image]], return_tensors="pt", return_row_col_info=True)
+
+        # Small image should not be tiled (1x1 grid), no thumbnail added
+        self.assertEqual(result.image_rows[0].item(), 1)
+        self.assertEqual(result.image_cols[0].item(), 1)
+        # Should have exactly 1 image in batch (no thumbnail)
+        self.assertEqual(result.pixel_values.shape[0], 1)
+
+    def test_large_image_no_tiling_smart_resize(self):
+        """Large image with tiling disabled should use smart resize, no thumbnail."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=False,
+            use_thumbnail=True,  # even if enabled, should not be used
+        )
+        # Create a large image (2048x2048)
+        large_image = Image.new("RGB", (2048, 2048), color="green")
+        result = image_processing([[large_image]], return_tensors="pt", return_row_col_info=True)
+
+        # With tiling disabled, should be 1 tile even for large images
+        self.assertEqual(result.image_rows[0].item(), 1)
+        self.assertEqual(result.image_cols[0].item(), 1)
+        # Should have exactly 1 image in batch (no thumbnail, smart resize only)
+        self.assertEqual(result.pixel_values.shape[0], 1)
+
+    def test_large_image_tiling_enabled_thumbnail_disabled(self):
+        """Large image with tiling enabled but thumbnail disabled should tile without thumbnail."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=False,
+            min_tiles=2,
+            max_tiles=10,
+            tile_size=512,
+        )
+        # Create a large image that will require tiling
+        large_image = Image.new("RGB", (2048, 2048), color="yellow")
+        result = image_processing([[large_image]], return_tensors="pt", return_row_col_info=True)
+
+        # Large image should be tiled into multiple tiles
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+        num_tiles = num_rows * num_cols
+        self.assertGreater(num_tiles, 1, "Large image should be tiled into multiple tiles")
+
+        # Count actual patches - with thumbnail disabled, should equal number of tiles
+        num_images_in_batch = result.pixel_values.shape[0]
+        self.assertEqual(
+            num_images_in_batch, num_tiles, "Number of patches should equal number of tiles (no thumbnail)"
+        )
+
+    def test_large_image_tiling_enabled_thumbnail_enabled(self):
+        """Large image with tiling and thumbnail enabled should tile AND add thumbnail."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=10,
+            tile_size=512,
+        )
+        # Create a large image that will require tiling
+        large_image = Image.new("RGB", (2048, 2048), color="purple")
+        result = image_processing([[large_image]], return_tensors="pt", return_row_col_info=True)
+
+        # Large image should be tiled into multiple tiles
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+        num_tiles = num_rows * num_cols
+        self.assertGreater(num_tiles, 1, "Large image should be tiled into multiple tiles")
+
+        # With thumbnail enabled, we should have tiles + 1 thumbnail
+        num_images_in_batch = result.pixel_values.shape[0]
+        self.assertEqual(num_images_in_batch, num_tiles + 1, "Number of patches should equal tiles + 1 (thumbnail)")
+
+    # ==================== Non-Square Aspect Ratio Tests ====================
+
+    def test_landscape_image_aspect_ratio(self):
+        """Test that landscape images (wider than tall) are processed correctly."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=10,
+            tile_size=512,
+        )
+        # Create a landscape image (1920x1080, ~16:9 aspect ratio)
+        landscape_image = Image.new("RGB", (1920, 1080), color="blue")
+        result = image_processing([[landscape_image]], return_tensors="pt", return_row_col_info=True)
+
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+
+        # Landscape image should have more columns than rows
+        self.assertGreaterEqual(num_cols, num_rows, "Landscape image should have cols >= rows")
+
+    def test_extreme_aspect_ratio_wide(self):
+        """Test extremely wide image (panorama-like)."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=10,
+            tile_size=512,
+        )
+        # Create an extremely wide image (3000x500)
+        wide_image = Image.new("RGB", (3000, 500), color="red")
+        result = image_processing([[wide_image]], return_tensors="pt", return_row_col_info=True)
+
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+
+        # Very wide image should have significantly more cols than rows
+        self.assertGreater(num_cols, num_rows, "Very wide image should have cols > rows")
+
+    def test_extreme_aspect_ratio_tall(self):
+        """Test extremely tall image."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=10,
+            tile_size=512,
+        )
+        # Create an extremely tall image (500x3000)
+        tall_image = Image.new("RGB", (500, 3000), color="yellow")
+        result = image_processing([[tall_image]], return_tensors="pt", return_row_col_info=True)
+
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+
+        # Very tall image should have significantly more rows than cols
+        self.assertGreater(num_rows, num_cols, "Very tall image should have rows > cols")
+
+    # ==================== Output Validation Tests ====================
+
+    def test_image_sizes_returned_with_row_col_info(self):
+        """Test that image_sizes is returned when return_row_col_info=True."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+        image = Image.new("RGB", (512, 256), color="green")
+        result = image_processing([[image]], return_tensors="pt", return_row_col_info=True)
+
+        # Check all row/col info is returned
+        self.assertIn("image_rows", result)
+        self.assertIn("image_cols", result)
+        self.assertIn("image_sizes", result)
+
+        # image_sizes should contain [height, width] for the resized image
+        image_sizes = result.image_sizes
+        self.assertIsInstance(image_sizes, torch.Tensor)
+        self.assertEqual(image_sizes.shape[0], 1)  # one sample
+        self.assertEqual(image_sizes.shape[1], 2)  # [height, width]
+
+    def test_output_consistency_across_formats(self):
+        """Test that outputs are consistent regardless of input format (PIL, numpy, torch)."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Create same image in different formats
+        pil_image = Image.new("RGB", (256, 256), color="white")
+        np_image = np.array(pil_image)
+        torch_image = torch.from_numpy(np_image).permute(2, 0, 1)
+
+        result_pil = image_processing([[pil_image]], return_tensors="pt")
+        result_np = image_processing([[np_image]], return_tensors="pt")
+        result_torch = image_processing([[torch_image]], return_tensors="pt")
+
+        # All should produce same shapes
+        self.assertEqual(result_pil.pixel_values.shape, result_np.pixel_values.shape)
+        self.assertEqual(result_pil.pixel_values.shape, result_torch.pixel_values.shape)
+        self.assertEqual(result_pil.spatial_shapes.tolist(), result_np.spatial_shapes.tolist())
+        self.assertEqual(result_pil.spatial_shapes.tolist(), result_torch.spatial_shapes.tolist())
+
+    # ==================== Multiple Images Per Sample Tests ====================
+
+    def test_multiple_images_per_sample(self):
+        """Test processing multiple images in a single sample: [[img1, img2, img3]]."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        img1 = Image.new("RGB", (256, 256), color="red")
+        img2 = Image.new("RGB", (256, 256), color="green")
+        img3 = Image.new("RGB", (256, 256), color="blue")
+
+        result = image_processing([[img1, img2, img3]], return_tensors="pt")
+
+        # Should have 3 images processed
+        self.assertEqual(result.pixel_values.shape[0], 3)
+        self.assertEqual(result.spatial_shapes.shape[0], 3)
+        self.assertEqual(result.pixel_attention_mask.shape[0], 3)
+
+    def test_mixed_image_counts_across_batch(self):
+        """Test batch with different number of images per sample: [[img1], [img2, img3]]."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        img1 = Image.new("RGB", (256, 256), color="red")
+        img2 = Image.new("RGB", (256, 256), color="green")
+        img3 = Image.new("RGB", (256, 256), color="blue")
+
+        # First sample has 1 image, second sample has 2 images
+        result = image_processing([[img1], [img2, img3]], return_tensors="pt")
+
+        # Total should be 3 images (1 + 2)
+        self.assertEqual(result.pixel_values.shape[0], 3)
+        self.assertEqual(result.spatial_shapes.shape[0], 3)
+
+    def test_multiple_images_different_sizes(self):
+        """Test multiple images per sample with different sizes."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        img_small = Image.new("RGB", (256, 256), color="red")
+        img_medium = Image.new("RGB", (512, 512), color="green")
+        img_large = Image.new("RGB", (768, 768), color="blue")
+
+        result = image_processing([[img_small, img_medium, img_large]], return_tensors="pt")
+
+        # Should have 3 images processed
+        self.assertEqual(result.pixel_values.shape[0], 3)
+        # All should have same max_num_patches due to padding
+        self.assertEqual(result.pixel_values.shape[1], image_processing.max_num_patches)
+
+    # ==================== Parameter Variations Tests ====================
+
+    def test_forced_grid_config_min_equals_max(self):
+        """Test forcing a specific grid configuration with min_tiles == max_tiles."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            min_tiles=4,
+            max_tiles=4,  # Force exactly 4 tiles
+            tile_size=512,
+            use_thumbnail=False,
+        )
+        # Large image that would normally get more tiles
+        wide_image = Image.new("RGB", (3000, 500), color="red")
+        result = image_processing([[wide_image]], return_tensors="pt", return_row_col_info=True)
+
+        num_rows = result.image_rows[0].item()
+        num_cols = result.image_cols[0].item()
+        num_tiles = num_rows * num_cols
+
+        # Should be exactly 4 tiles
+        self.assertEqual(num_tiles, 4, "Should have exactly 4 tiles when min_tiles == max_tiles == 4")
+
+    # ==================== Input Validation Tests ====================
+
+    def test_min_tiles_greater_than_max_tiles_raises_error(self):
+        """Test that min_tiles > max_tiles raises ValueError."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            min_tiles=10,
+            max_tiles=2,  # Invalid: min > max
+        )
+        image = Image.new("RGB", (1024, 1024), color="red")
+
+        with self.assertRaises(ValueError) as context:
+            image_processing([[image]], return_tensors="pt")
+
+        self.assertIn("min_tiles", str(context.exception).lower())
+
+    # ==================== Edge Case Images Tests ====================
+
+    def test_very_small_image(self):
+        """Test image smaller than encoder_patch_size."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=False,
+            encoder_patch_size=16,
+        )
+        # Image smaller than patch size
+        tiny_image = Image.new("RGB", (8, 8), color="red")
+        result = image_processing([[tiny_image]], return_tensors="pt")
+
+        # Should still process without error
+        self.assertIn("pixel_values", result)
+        self.assertEqual(result.pixel_values.dim(), 3)
+
+    def test_grayscale_image(self):
+        """Test that grayscale (1-channel) images are converted to RGB."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Create grayscale image
+        grayscale_image = Image.new("L", (256, 256), color=128)
+        result = image_processing([[grayscale_image]], return_tensors="pt")
+
+        # Should process and output 3 channels (converted to RGB)
+        self.assertIn("pixel_values", result)
+        # pixel_values shape is (batch, num_patches, patch_size^2 * 3)
+        expected_patch_dim = 3 * image_processing.encoder_patch_size**2
+        self.assertEqual(result.pixel_values.shape[2], expected_patch_dim)
+
+    def test_rgba_4_channel_image(self):
+        """Test that RGBA (4-channel) images are converted to RGB."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Create RGBA image with alpha channel
+        rgba_image = Image.new("RGBA", (256, 256), color=(255, 0, 0, 128))
+        result = image_processing([[rgba_image]], return_tensors="pt", do_convert_rgb=True)
+
+        # Should process and output 3 channels (alpha dropped)
+        self.assertIn("pixel_values", result)
+        expected_patch_dim = 3 * image_processing.encoder_patch_size**2
+        self.assertEqual(result.pixel_values.shape[2], expected_patch_dim)
+
+    def test_numpy_4_channel_rgba(self):
+        """Test actual 4-channel numpy array input - convert to PIL for RGB conversion."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Create 4-channel numpy array (RGBA) and convert to PIL Image for RGB conversion
+        rgba_np = np.random.randint(0, 255, (256, 256, 4), dtype=np.uint8)
+        rgba_pil = Image.fromarray(rgba_np, mode="RGBA")
+        result = image_processing([[rgba_pil]], return_tensors="pt", do_convert_rgb=True)
+
+        # Should convert to 3 channels
+        self.assertIn("pixel_values", result)
+        expected_patch_dim = 3 * image_processing.encoder_patch_size**2
+        self.assertEqual(result.pixel_values.shape[2], expected_patch_dim)
+
+    def test_single_pixel_image(self):
+        """Test 1x1 pixel image (extreme edge case)."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        single_pixel = Image.new("RGB", (1, 1), color="blue")
+        result = image_processing([[single_pixel]], return_tensors="pt")
+
+        # Should process without error
+        self.assertIn("pixel_values", result)
+
+    # ==================== Helper Function Unit Tests ====================
+
+    def test_round_by_factor(self):
+        """Test round_by_factor function."""
+        # Exact multiples should return themselves
+        self.assertEqual(round_by_factor(32, 16), 32)
+        self.assertEqual(round_by_factor(64, 16), 64)
+
+        # Values should round to nearest multiple
+        self.assertEqual(round_by_factor(30, 16), 32)  # 30 -> 32 (closer to 32 than 16)
+        self.assertEqual(round_by_factor(20, 16), 16)  # 20 -> 16 (closer to 16 than 32)
+        self.assertEqual(round_by_factor(24, 16), 32)  # 24 -> 32 (equidistant, rounds up)
+
+        # Test with different factors
+        self.assertEqual(round_by_factor(100, 32), 96)  # 100 -> 96
+        self.assertEqual(round_by_factor(50, 32), 64)  # 50 -> 64
+
+        # Test with factor of 1
+        self.assertEqual(round_by_factor(17, 1), 17)
+
+    def test_is_image_too_large_small_image(self):
+        """Test _is_image_too_large with small image."""
+        image_processing = self.image_processing_classes["torchvision"](
+            max_image_tokens=256,
+            encoder_patch_size=16,
+            downsample_factor=2,
+            max_pixels_tolerance=2.0,
+        )
+
+        is_large = image_processing._is_image_too_large(
+            height=512,
+            width=512,
+            max_image_tokens=256,
+            encoder_patch_size=16,
+            downsample_factor=2,
+            max_pixels_tolerance=2.0,
+        )
+        self.assertFalse(is_large)
+
+    def test_is_image_too_large_large_image(self):
+        """Test _is_image_too_large with large image."""
+        image_processing = self.image_processing_classes["torchvision"](
+            max_image_tokens=256,
+            encoder_patch_size=16,
+            downsample_factor=2,
+            max_pixels_tolerance=1.0,
+        )
+
+        is_large = image_processing._is_image_too_large(
+            height=565,
+            width=565,
+            max_image_tokens=256,
+            encoder_patch_size=16,
+            downsample_factor=2,
+            max_pixels_tolerance=1.0,
+        )
+        self.assertTrue(is_large)
+
+    # ==================== Batch Processing Tests ====================
+
+    def test_batch_mixed_image_sizes(self):
+        """Test batch processing with different image sizes requiring different processing paths."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Create images with significantly different sizes
+        small_image = Image.new("RGB", (256, 256), color="red")
+        medium_image = Image.new("RGB", (512, 512), color="green")
+        large_image = Image.new("RGB", (1024, 1024), color="blue")
+
+        # Process as batch
+        result = image_processing([[small_image], [medium_image], [large_image]], return_tensors="pt")
+
+        # All should be processed and padded to same size
+        self.assertEqual(result.pixel_values.shape[0], 3)
+        # All should have same max_num_patches
+        self.assertEqual(result.pixel_values.shape[1], image_processing.max_num_patches)
+        # Patch dimension should be patch_size^2 * 3 channels
+        expected_patch_dim = 3 * image_processing.encoder_patch_size**2
+        self.assertEqual(result.pixel_values.shape[2], expected_patch_dim)
+
+        # Spatial shapes should all be square (equal height and width)
+        shapes = result.spatial_shapes.tolist()
+        for shape in shapes:
+            self.assertEqual(shape[0], shape[1], "Square images should have equal height and width")
+
+        # pixel_attention_mask should have correct shape
+        self.assertEqual(result.pixel_attention_mask.shape[0], 3)
+        self.assertEqual(result.pixel_attention_mask.shape[1], image_processing.max_num_patches)
+
+    def test_batch_mixed_aspect_ratios(self):
+        """Test batch with mixed aspect ratios."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        square = Image.new("RGB", (512, 512), color="red")
+        landscape = Image.new("RGB", (1024, 512), color="green")
+        portrait = Image.new("RGB", (512, 1024), color="blue")
+
+        result = image_processing([[square], [landscape], [portrait]], return_tensors="pt")
+
+        # All should be processed
+        self.assertEqual(result.pixel_values.shape[0], 3)
+        self.assertEqual(result.spatial_shapes.shape[0], 3)
+
+        # Spatial shapes should reflect aspect ratios: [height, width]
+        shapes = result.spatial_shapes.tolist()
+        square_shape, landscape_shape, portrait_shape = shapes
+
+        # Square: height == width
+        self.assertEqual(square_shape[0], square_shape[1], "Square image should have equal spatial dimensions")
+
+        # Landscape: width > height
+        self.assertGreater(landscape_shape[1], landscape_shape[0], "Landscape image should have width > height")
+
+        # Portrait: height > width
+        self.assertGreater(portrait_shape[0], portrait_shape[1], "Portrait image should have height > width")
+
+        # pixel_attention_mask should match batch size and max_num_patches
+        self.assertEqual(result.pixel_attention_mask.shape[0], 3)
+        self.assertEqual(result.pixel_attention_mask.shape[1], image_processing.max_num_patches)
+
+    def test_disable_grouping_single_image(self):
+        """Test disable_grouping parameter with single image."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        image = Image.new("RGB", (512, 512), color="purple")
+
+        # Process with and without disable_grouping
+        result_grouped = image_processing([[image]], return_tensors="pt", disable_grouping=False)
+        result_ungrouped = image_processing([[image]], return_tensors="pt", disable_grouping=True)
+
+        # Both should produce all expected output keys
+        for result in [result_grouped, result_ungrouped]:
+            self.assertIn("pixel_values", result)
+            self.assertIn("spatial_shapes", result)
+            self.assertIn("pixel_attention_mask", result)
+
+        # Both should have same output shapes for single image
+        self.assertEqual(result_grouped.pixel_values.shape, result_ungrouped.pixel_values.shape)
+        self.assertEqual(result_grouped.spatial_shapes.shape, result_ungrouped.spatial_shapes.shape)
+        self.assertEqual(result_grouped.pixel_attention_mask.shape, result_ungrouped.pixel_attention_mask.shape)
+
+        # Verify specific shapes
+        self.assertEqual(result_ungrouped.pixel_values.shape[0], 1)
+        self.assertEqual(result_ungrouped.pixel_values.shape[1], image_processing.max_num_patches)
+        expected_patch_dim = 3 * image_processing.encoder_patch_size**2
+        self.assertEqual(result_ungrouped.pixel_values.shape[2], expected_patch_dim)
+
+    def test_disable_grouping_batch(self):
+        """Test disable_grouping parameter with batch of images."""
+        image_processing = self.image_processing_classes["torchvision"](do_image_splitting=False)
+
+        # Images of same size - normally would be grouped
+        img1 = Image.new("RGB", (256, 256), color="red")
+        img2 = Image.new("RGB", (256, 256), color="green")
+        img3 = Image.new("RGB", (256, 256), color="blue")
+
+        # Process with disable_grouping=True
+        result = image_processing([[img1], [img2], [img3]], return_tensors="pt", disable_grouping=True)
+
+        # Should produce valid output for all images
+        self.assertEqual(result.pixel_values.shape[0], 3)
+
+    def test_batch_with_tiling(self):
+        """Test batch processing when some images need tiling."""
+        image_processing = self.image_processing_classes["torchvision"](
+            do_image_splitting=True,
+            use_thumbnail=True,
+            min_tiles=2,
+            max_tiles=4,
+            tile_size=512,
+        )
+
+        # Small image (no tiling needed) and large image (will be tiled)
+        small = Image.new("RGB", (256, 256), color="red")
+        large = Image.new("RGB", (1024, 1024), color="blue")  # 2x2 tiles at 512
+
+        result = image_processing([[small], [large]], return_tensors="pt", return_row_col_info=True)
+
+        # Calculate tiles for each image
+        small_tiles = result.image_rows[0].item() * result.image_cols[0].item()
+        large_tiles = result.image_rows[1].item() * result.image_cols[1].item()
+
+        # Small image: single tile (no splitting needed)
+        self.assertEqual(small_tiles, 1, "Small 256x256 image should have 1 tile (no splitting)")
+
+        # Large image: 2x2 = 4 tiles for 1024x1024 with tile_size=512
+        self.assertEqual(large_tiles, 4, "Large 1536x1536 image should have 4 tiles (2x2)")
+
+        # Total images: small (1) + large tiles (4) + thumbnail for large (1) = 6
+        # Thumbnail is only added when there's more than 1 tile
+        expected_total = 1 + 4 + 1  # small + large_tiles + large_thumbnail
+        self.assertEqual(result.pixel_values.shape[0], expected_total)
+        self.assertEqual(result.spatial_shapes.shape[0], expected_total)
+        self.assertEqual(result.pixel_attention_mask.shape[0], expected_total)
+
+    def test_batch_tiling(self):
+        """Test that patches from different images don't get mixed when batch processing with tiling.
+
+        This test verifies that when processing a batch of images with tiling enabled,
+        patches from image 0 don't end up in image 1's output and vice versa.
+        This was a bug caused by incorrect permute/reshape operations in crop_image_to_patches.
+        """
+        for use_thumbnail in [False, True]:
+            with self.subTest(use_thumbnail=use_thumbnail):
+                image_processing = self.image_processing_classes["torchvision"](
+                    do_image_splitting=True,
+                    use_thumbnail=use_thumbnail,
+                    min_tiles=2,
+                    max_tiles=4,
+                    tile_size=512,
+                )
+
+                # Create two large images with completely different solid colors
+                # Red image: RGB = (255, 0, 0)
+                # Blue image: RGB = (0, 0, 255)
+                red_image = Image.new("RGB", (1024, 1024), color=(255, 0, 0))
+                blue_image = Image.new("RGB", (1024, 1024), color=(0, 0, 255))
+
+                result = image_processing(
+                    [[red_image], [blue_image]],
+                    return_tensors="pt",
+                    return_row_col_info=True,
+                    do_rescale=False,  # Keep original pixel values for easier verification
+                    do_normalize=False,
+                )
+
+                # Each 1024x1024 image should be split into 2x2 = 4 tiles
+                red_tiles = result.image_rows[0].item() * result.image_cols[0].item()
+                blue_tiles = result.image_rows[1].item() * result.image_cols[1].item()
+                self.assertEqual(red_tiles, 4)
+                self.assertEqual(blue_tiles, 4)
+
+                # Calculate expected total patches
+                # Without thumbnail: 4 + 4 = 8
+                # With thumbnail: (4 + 1) + (4 + 1) = 10
+                thumb_count = 1 if use_thumbnail else 0
+                expected_total = (red_tiles + thumb_count) + (blue_tiles + thumb_count)
+                self.assertEqual(result.pixel_values.shape[0], expected_total)
+
+                pixel_values = result.pixel_values
+                patch_size = image_processing.encoder_patch_size
+                patches_per_image = red_tiles + thumb_count
+
+                # Check red image patches (and thumbnail if enabled)
+                # All should have high red, zero blue
+                for i in range(patches_per_image):
+                    first_patch = pixel_values[i][0].view(3, patch_size, patch_size)
+                    red_mean = first_patch[0].float().mean().item()
+                    blue_mean = first_patch[2].float().mean().item()
+                    patch_type = "thumbnail" if use_thumbnail and i == red_tiles else f"patch {i}"
+                    self.assertGreater(
+                        red_mean,
+                        blue_mean,
+                        f"Red image {patch_type} has more blue than red - patches may be interleaved",
+                    )
+
+                # Check blue image patches (and thumbnail if enabled)
+                # All should have high blue, zero red
+                for i in range(patches_per_image, 2 * patches_per_image):
+                    first_patch = pixel_values[i][0].view(3, patch_size, patch_size)
+                    red_mean = first_patch[0].float().mean().item()
+                    blue_mean = first_patch[2].float().mean().item()
+                    local_idx = i - patches_per_image
+                    patch_type = "thumbnail" if use_thumbnail and local_idx == blue_tiles else f"patch {local_idx}"
+                    self.assertGreater(
+                        blue_mean,
+                        red_mean,
+                        f"Blue image {patch_type} has more red than blue - patches may be interleaved",
+                    )

tests/models/lfm2_vl/test_modeling_lfm2_vl.py

@@ -0,0 +1,385 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the LFM2-VL model."""
+
+import math
+import unittest
+from io import BytesIO
+
+import pytest
+import requests
+
+from transformers import AutoProcessor, is_torch_available
+from transformers.models.lfm2_vl.modeling_lfm2_vl import Lfm2VlForConditionalGeneration
+from transformers.testing_utils import (
+    Expectations,
+    cleanup,
+    require_deterministic_for_xpu,
+    require_torch,
+    require_torch_accelerator,
+    slow,
+    torch_device,
+)
+from transformers.utils.import_utils import is_vision_available
+
+from ...causal_lm_tester import CausalLMModelTester
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_vision_available():
+    from PIL import Image
+
+if is_torch_available():
+    import torch
+
+    from transformers import Lfm2VlConfig, Lfm2VlForConditionalGeneration, Lfm2VlModel
+
+
+class Lfm2VlModelTester(CausalLMModelTester):
+    if is_torch_available():
+        config_class = Lfm2VlConfig
+        base_model_class = Lfm2VlModel
+        causal_lm_class = Lfm2VlForConditionalGeneration
+
+    def __init__(
+        self,
+        parent,
+        is_training=True,
+        batch_size=2,
+        scale_factor=2,
+        num_images=2,
+        vision_config={
+            "hidden_size": 32,
+            "intermediate_size": 37,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 2,
+            "num_channels": 3,
+            "num_patches": 16,
+            "patch_size": 4,
+            "hidden_act": "gelu_pytorch_tanh",
+            "layer_norm_eps": 1e-6,
+            "attention_dropout": 0.0,
+        },
+        text_config={
+            "vocab_size": 100,
+            "hidden_size": 32,
+            "intermediate_size": 37,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "num_key_value_heads": 2,
+            "max_position_embeddings": 100,
+            "pad_token_id": 0,
+            "bos_token_id": 1,
+            "eos_token_id": 2,
+            "tie_word_embeddings": True,
+            "rope_theta": 1000000.0,
+            "conv_bias": False,
+            "conv_L_cache": 3,
+            "block_multiple_of": 2,
+            "full_attn_idxs": [0],
+        },
+        image_token_id=4,
+        downsample_factor=4,
+        projector_hidden_size=32,
+    ):
+        super().__init__(parent)
+        self.vision_config = vision_config
+        self.text_config = text_config
+        self.image_token_id = image_token_id
+        self.is_training = is_training
+        self.batch_size = batch_size
+        self.scale_factor = scale_factor
+        self.num_images = num_images
+        self.downsample_factor = downsample_factor
+        self.projector_hidden_size = projector_hidden_size
+        self.image_seq_length = 4
+
+    def get_config(self):
+        return Lfm2VlConfig(
+            vision_config=self.vision_config,
+            text_config=self.text_config,
+            image_token_id=self.image_token_id,
+            downsample_factor=self.downsample_factor,
+            projector_hidden_size=self.projector_hidden_size,
+        )
+
+    def prepare_config_and_inputs(self):
+        # Create dummy pixel values: [num_images, num_patches, channels * patch_size^2]
+        patch_size = self.vision_config["patch_size"]
+        pixel_values = floats_tensor([self.num_images, 64, 3 * patch_size * patch_size])
+
+        # Spatial shapes: one (height_patches, width_patches) per image
+        patches = int(math.sqrt(64))
+        spatial_shapes = torch.tensor([[patches, patches]] * self.num_images, dtype=torch.long, device=torch_device)
+
+        # Pixel attention mask: mark all patches as valid (no padding)
+        pixel_attention_mask = torch.ones((self.num_images, 64), dtype=torch.long, device=torch_device)
+        config = self.get_config()
+        return config, pixel_values, spatial_shapes, pixel_attention_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, spatial_shapes, pixel_attention_mask = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 2) + 1
+
+        # For simplicity just set the last n tokens to the image token
+        input_ids[input_ids == self.image_token_id] = self.text_config["pad_token_id"]
+        input_ids[:, : self.image_seq_length] = self.image_token_id
+
+        attention_mask = input_ids.ne(1).to(torch_device)
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "spatial_shapes": spatial_shapes,
+            "pixel_attention_mask": pixel_attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Lfm2VlModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (Lfm2VlModel, Lfm2VlForConditionalGeneration) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": Lfm2VlModel,
+            "text-generation": Lfm2VlForConditionalGeneration,
+        }
+        if is_torch_available()
+        else {}
+    )
+
+    model_tester_class = Lfm2VlModelTester
+    _is_composite = True
+    test_torch_exportable = False
+
+    def setUp(self):
+        self.model_tester = Lfm2VlModelTester(self)
+        common_properties = ["image_token_id", "projector_hidden_size"]
+        self.config_tester = ConfigTester(
+            self, config_class=Lfm2VlConfig, has_text_modality=False, common_properties=common_properties
+        )
+
+    def _get_conv_state_shape(self, batch_size: int, config):
+        return (batch_size, config.hidden_size, config.conv_L_cache)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(
+        "Lfm2 backbone alternates between attention and conv layers, so attention are only returned for attention layers"
+    )
+    def test_attention_outputs(self):
+        pass
+
+    @unittest.skip(
+        "Lfm2 backbone has a special cache format which is not compatible with compile as it has static address for conv cache"
+    )
+    @pytest.mark.torch_compile_test
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing(self):
+        super().test_training_gradient_checkpointing()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        super().test_training_gradient_checkpointing_use_reentrant_false()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        super().test_training_gradient_checkpointing_use_reentrant_true()
+
+
+@require_torch_accelerator
+@slow
+class Lfm2VlForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("LiquidAI/LFM2-VL-1.6B")
+        self.processor.tokenizer.padding_side = "left"
+        self.image = Image.open(
+            requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw
+        )
+        self.image2 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    @require_deterministic_for_xpu
+    def test_integration_test(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = "In this image, we see two cats sleeping on a pink blanket. There are also two remote controls on the blanket.\n\n\n\n"
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    @require_deterministic_for_xpu
+    def test_integration_test_high_resolution(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image2
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = (
+            "In this image, we see the Statue of Liberty, standing tall on its pedestal. The statue is made of metal,"
+        )
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    @require_deterministic_for_xpu
+    def test_integration_test_batched(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2-VL-450M",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = ["<image>In this image, we see", "<image>In this image, we see a cat"]
+        images = [[self.image2], [self.image]]
+        inputs = self.processor(text=text, images=images, return_tensors="pt", padding=True)
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = [
+            "In this image, we see a panoramic view of the New York City skyline. The iconic Statics and the New York",
+            "In this image, we see a cat that is lying on its side on a cat bed.",
+        ]
+        self.assertListEqual(generated_texts, expected_generated_text)
+
+
+@require_torch_accelerator
+@slow
+class Lfm2_5VlForConditionalGenerationIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.processor = AutoProcessor.from_pretrained("LiquidAI/LFM2.5-VL-1.6B")
+        self.processor.tokenizer.padding_side = "left"
+        self.image = Image.open(
+            requests.get("http://images.cocodataset.org/val2017/000000039769.jpg", stream=True).raw
+        )
+        self.image2 = Image.open(
+            BytesIO(
+                requests.get(
+                    "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
+                ).content
+            )
+        )
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    @require_deterministic_for_xpu
+    def test_integration_test(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2.5-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = (
+            "In this image, we see two cats lying on a pink blanket. One cat is a tabby, and the other is a"
+        )
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    @require_deterministic_for_xpu
+    def test_integration_test_high_resolution(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2.5-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = "<image>In this image, we see"
+        images = self.image2
+        inputs = self.processor(text=text, images=images, return_tensors="pt")
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = "In this image, we see the Statue of Liberty, an iconic symbol of freedom and democracy. It stands on Liberty Island in"
+        self.assertEqual(generated_texts[0], expected_generated_text)
+
+    @require_deterministic_for_xpu
+    def test_integration_test_batched(self):
+        model = Lfm2VlForConditionalGeneration.from_pretrained(
+            "LiquidAI/LFM2.5-VL-1.6B",
+            dtype=torch.bfloat16,
+            device_map="auto",
+        )
+
+        # Create inputs
+        text = ["<image>In this image, we see", "<image>In this image, we see"]
+        images = [[self.image2], [self.image]]
+        inputs = self.processor(text=text, images=images, return_tensors="pt", padding=True)
+        inputs.to(device=torch_device, dtype=torch.bfloat16)
+
+        generated_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        generated_texts = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+        expected_generated_text = Expectations(
+            {
+                (None, None): [
+                    "In this image, we see the Statue of Liberty, an iconic symbol of freedom and democracy. It stands on Liberty Island in",
+                    "In this image, we see two cats lying on a pink blanket. One cat is a tabby, and the other is a",
+                ],
+                ("xpu", 5): [
+                    "In this image, we see the Statue of Liberty, an iconic symbol of freedom and democracy. It stands tall on a small",
+                    "In this image, we see two cats lying on a pink blanket. One cat is a tabby, and the other is a",
+                ],
+            }
+        ).get_expectation()
+        self.assertListEqual(generated_texts, expected_generated_text)

tests/models/lfm2_vl/test_processing_lfm2_vl.py

@@ -0,0 +1,599 @@
+# Copyright 2025 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import unittest
+
+import numpy as np
+
+from transformers import Lfm2VlProcessor
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torchvision_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from PIL import Image
+
+    if is_torchvision_available():
+        pass
+
+
+@require_torch
+@require_vision
+class Lfm2VlProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = Lfm2VlProcessor
+
+    @classmethod
+    def _setup_image_processor(cls):
+        image_processor_class = cls._get_component_class_from_processor("image_processor")
+        return image_processor_class(
+            tile_size=14,
+            min_image_tokens=2,
+            max_image_tokens=10,
+            encoder_patch_size=2,
+            do_image_splitting=False,
+        )
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
+        processor_kwargs = cls.prepare_processor_dict()
+        return tokenizer_class.from_pretrained("LiquidAI/LFM2-VL-1.6B", **processor_kwargs)
+
+    @classmethod
+    def _setup_test_attributes(cls, processor):
+        # Create images with different sizes
+        cls.small_image = Image.new("RGB", (256, 256))
+        cls.large_image = Image.new("RGB", (512, 1024))
+        cls.high_res_image = Image.new("RGB", (1024, 1024))
+        cls.bos_token = processor.tokenizer.bos_token
+        cls.image_token = processor.image_token
+
+        cls.bos_token_id = processor.tokenizer.convert_tokens_to_ids(cls.bos_token)
+        cls.image_token_id = processor.image_token_id
+        cls.image_start_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_start_token)
+        cls.image_end_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_end_token)
+        cls.padding_token_id = processor.tokenizer.pad_token_id
+        cls.image_thumbnail_token_id = processor.tokenizer.convert_tokens_to_ids(processor.image_thumbnail_token)
+
+    @staticmethod
+    def prepare_processor_dict():
+        chat_template = (
+            "{{bos_token}}{% for message in messages %}"
+            "{{'<|im_start|>' + message['role'] + '\n'}}"
+            "{% if message['content'] is string %}"
+            "{{ message['content'] }}"
+            "{% else %}"
+            "{% for content in message['content'] %}"
+            "{% if content['type'] == 'image' %}"
+            "{{ '<image>' }}"
+            "{% elif content['type'] == 'text' %}"
+            "{{ content['text'] }}"
+            "{% endif %}"
+            "{% endfor %}"
+            "{% endif %}"
+            "{{'<|im_end|>\n'}}"
+            "{% endfor %}"
+            "{% if add_generation_prompt %}"
+            "{{'<|im_start|>assistant\n' }}"
+            "{% endif %}"
+        )
+        return {"chat_template": chat_template}
+
+    @unittest.skip("Lfm2VlProcessor adds special tokens to the text")
+    def test_tokenizer_defaults(self):
+        pass
+
+    # Override as Lfm2VL needs images/video to be an explicitly nested batch
+    def prepare_image_inputs(self, batch_size=None):
+        """This function prepares a list of PIL images for testing"""
+        images = super().prepare_image_inputs(batch_size)
+        if isinstance(images, (list, tuple)):
+            images = [[image] for image in images]
+        return images
+
+    def get_split_image_expected_tokens(self, processor, image_rows, image_cols, add_thumbnail, image_seq_len):
+        text_split_images = [self.image_start_token_id]
+        num_patches_tile = processor.image_processor.tile_size // processor.image_processor.encoder_patch_size
+        tile_seq_len = math.ceil(num_patches_tile / processor.image_processor.downsample_factor) ** 2
+        for n_h in range(image_rows):
+            for n_w in range(image_cols):
+                text_split_images += (
+                    processor.tokenizer(f"<|img_row_{n_h + 1}_col_{n_w + 1}|>", add_special_tokens=False)["input_ids"]
+                    + [self.image_token_id] * tile_seq_len
+                )
+        if add_thumbnail:
+            text_split_images += [self.image_thumbnail_token_id] + [self.image_token_id] * image_seq_len
+        text_split_images += [self.image_end_token_id]
+        return text_split_images
+
+    def test_process_interleaved_images_prompts_no_image_splitting_single_image(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+        image_str = "<image>"
+
+        # Test that a single image is processed correctly
+        inputs = processor(images=self.small_image, text=image_str)
+        encoder_feature_dims = (
+            3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_values"]).shape,
+            (1, processor.image_processor.max_num_patches, encoder_feature_dims),
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_attention_mask"]).shape, (1, processor.image_processor.max_num_patches)
+        )
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6]])
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_no_image_splitting_single_image_with_text(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+        inputs = processor(text=text, images=self.small_image)
+
+        # fmt: off
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        expected_input_ids = [[self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id] + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        encoder_feature_dims = 3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        self.assertEqual(np.array(inputs["pixel_values"]).shape, (1, processor.image_processor.max_num_patches, encoder_feature_dims))
+        self.assertEqual(np.array(inputs["pixel_attention_mask"]).shape, (1, processor.image_processor.max_num_patches))
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6]])
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_no_image_splitting_multiple_images(self):
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "In this image, we see"
+
+        text = [
+            image_str + text_str_1,
+            image_str + image_str + text_str_2,
+        ]
+        images = [[self.small_image], [self.small_image, self.small_image]]
+
+        inputs = processor(text=text, images=images, padding=True)
+
+        tokenized_sentence_1 = processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = processor.tokenizer(text_str_2, add_special_tokens=False)
+        image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+        expected_input_ids_1 = image_tokens + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = 2 * image_tokens + tokenized_sentence_2["input_ids"]
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [self.padding_token_id] * pad_len + expected_input_ids_1
+
+        self.assertEqual(inputs["input_ids"], [padded_expected_input_ids_1, expected_input_ids_2])
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)],
+        )
+        encoder_feature_dims = (
+            3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_values"]).shape,
+            (3, processor.image_processor.max_num_patches, encoder_feature_dims),
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_attention_mask"]).shape, (3, processor.image_processor.max_num_patches)
+        )
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), [[6, 6], [6, 6], [6, 6]])
+
+    def test_process_interleaved_images_prompts_image_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str_1 = "In this image, we see"
+        text_str_2 = "bla, bla"
+
+        text = [image_str + text_str_1, text_str_2 + image_str + image_str]
+        images = [[self.small_image], [self.high_res_image, self.high_res_image]]
+
+        inputs = processor(
+            text=text,
+            images=images,
+            padding=True,
+            padding_side="left",
+            max_pixels_tolerance=2.0,
+            use_thumbnail=True,
+            do_image_splitting=True,
+        )
+
+        tokenized_sentence_1 = processor.tokenizer(text_str_1, add_special_tokens=False)
+        tokenized_sentence_2 = processor.tokenizer(text_str_2, add_special_tokens=False)
+
+        small_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        large_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        high_res_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+
+        expected_input_ids_1 = small_image_tokens + tokenized_sentence_1["input_ids"]
+        expected_input_ids_2 = tokenized_sentence_2["input_ids"] + large_image_tokens + high_res_image_tokens
+        # Pad the first input to match the second input
+        pad_len = len(expected_input_ids_2) - len(expected_input_ids_1)
+        padded_expected_input_ids_1 = [self.padding_token_id] * pad_len + expected_input_ids_1
+
+        self.assertEqual(inputs["input_ids"][0], padded_expected_input_ids_1)
+        self.assertEqual(inputs["input_ids"][1], expected_input_ids_2)
+        self.assertEqual(
+            inputs["attention_mask"],
+            [[0] * pad_len + [1] * len(expected_input_ids_1), [1] * len(expected_input_ids_2)],
+        )
+        self.assertEqual(np.array(inputs["pixel_values"]).shape, (30, 49, 12))
+        self.assertEqual(np.array(inputs["pixel_attention_mask"]).shape, (30, 49))
+        self.assertListEqual(inputs["spatial_shapes"].tolist(), ([[7, 7]] * 9 + [[6, 6]]) * 3)
+
+    def test_add_special_tokens_processor_image_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = text_str + image_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.high_res_image, add_special_tokens=False, do_image_splitting=True)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        split_high_res_image_tokens = self.get_split_image_expected_tokens(processor, 3, 3, True, 9)
+        expected_input_ids = [tokenized_sentence["input_ids"] + split_high_res_image_tokens]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_add_special_tokens_processor_image_splitting_large_image(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = text_str + image_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.large_image, add_special_tokens=False, max_pixels_tolerance=2.0, do_image_splitting=True)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        large_image_tokens = self.get_split_image_expected_tokens(processor, 4, 2, True, 8)
+        expected_input_ids = [tokenized_sentence["input_ids"] + large_image_tokens]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_add_special_tokens_processor_image_no_splitting(self):
+        processor = self.get_processor()
+
+        image_str = "<image>"
+        text_str = "In this image, we see"
+        text = image_str + text_str
+
+        # fmt: off
+        inputs = processor(text=text, images=self.high_res_image, add_special_tokens=False, use_image_special_tokens=True, do_image_splitting=False)
+        tokenized_sentence = processor.tokenizer(text_str, add_special_tokens=False)
+        split_high_res_image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+        expected_input_ids = [split_high_res_image_tokens + tokenized_sentence["input_ids"]]
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        # fmt: on
+
+    def test_process_interleaved_images_prompts_image_error(self):
+        processor = self.get_processor()
+
+        text = [
+            "This is a test sentence.",
+            "In this other sentence we try some good things",
+        ]
+        images = [[self.small_image], [self.large_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [[self.small_image], []]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        text = [
+            "This is a test sentence.<image>",
+            "In this other sentence we try some good things<image>",
+        ]
+        images = [[self.small_image], [self.large_image, self.high_res_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [[], [self.large_image]]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [self.small_image, self.large_image, self.high_res_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+        images = [self.small_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        text = [
+            "This is a test sentence.",
+            "In this other sentence we try some good things<image>",
+        ]
+        images = [[self.small_image], []]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        images = [[], [self.large_image]]
+        processor(text=text, images=images, padding=True)
+
+        images = [self.small_image, self.large_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+        images = [self.small_image]
+        with self.assertRaises(ValueError):
+            processor(text=text, images=images, padding=True)
+
+    def test_apply_chat_template(self):
+        # Message contains content which a mix of lists with images and image urls and string
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What do these images show?"},
+                    {"type": "image"},
+                    {"type": "image"},
+                ],
+            },
+            {
+                "role": "assistant",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": "The first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.",
+                    }
+                ],
+            },
+            {"role": "user", "content": [{"type": "text", "text": "And who is that?"}]},
+        ]
+        processor = self.get_processor()
+        # Make short sequence length to test that the fake tokens are added correctly
+        rendered = processor.apply_chat_template(messages, add_generation_prompt=True)
+
+        expected_rendered = (
+            "<|startoftext|><|im_start|>user\nWhat do these images show?<image><image><|im_end|>\n"
+            "<|im_start|>assistant\nThe first image shows the statue of Liberty in New York. The second image picture depicts Idefix, the dog of Obelix in Asterix and Obelix.<|im_end|>\n"
+            "<|im_start|>user\nAnd who is that?<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        self.assertEqual(rendered, expected_rendered)
+
+    def test_text_only_inference(self):
+        """Test that the processor works correctly with text-only input."""
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        text = "This is a simple text without images."
+        inputs = processor(text=text)
+
+        tokenized_sentence = processor.tokenizer(text, add_special_tokens=False)
+        expected_input_ids = [tokenized_sentence["input_ids"]]
+
+        self.assertEqual(inputs["input_ids"], expected_input_ids)
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids[0])])
+        self.assertTrue("pixel_values" not in inputs)
+        self.assertTrue("pixel_attention_mask" not in inputs)
+
+        # Test batch of texts without image tokens
+        texts = ["First text.", "Second piece of text."]
+        batch_inputs = processor(text=texts, padding=True)
+
+        tokenized_1 = processor.tokenizer(texts[0], add_special_tokens=False)
+        tokenized_2 = processor.tokenizer(texts[1], add_special_tokens=False)
+
+        expected_1 = tokenized_1["input_ids"]
+        expected_2 = tokenized_2["input_ids"]
+
+        # Pad the shorter sequence
+        pad_len = len(expected_2) - len(expected_1)
+        if pad_len > 0:
+            padded_expected_1 = [self.padding_token_id] * pad_len + expected_1
+            expected_attention_1 = [0] * pad_len + [1] * len(expected_1)
+            self.assertEqual(batch_inputs["input_ids"], [padded_expected_1, expected_2])
+            self.assertEqual(batch_inputs["attention_mask"], [expected_attention_1, [1] * len(expected_2)])
+        else:
+            pad_len = -pad_len
+            padded_expected_2 = [self.padding_token_id] * pad_len + expected_2
+            expected_attention_2 = [0] * pad_len + [1] * len(expected_2)
+            self.assertEqual(batch_inputs["input_ids"], [expected_1, padded_expected_2])
+            self.assertEqual(batch_inputs["attention_mask"], [[1] * len(expected_1), expected_attention_2])
+
+    def test_missing_images_error(self):
+        """Test that appropriate error is raised when images are referenced but not provided."""
+        processor = self.get_processor()
+
+        # Test single text with image token but no image
+        text = "Let me show you this image: <image> What do you think?"
+        with self.assertRaises(ValueError) as context:
+            processor(text=text)
+        self.assertTrue("We detected 1 tokens in the text but no images were passed" in str(context.exception))
+
+        # Test batch with image tokens but no images
+        texts = [
+            "First text with <image> token.",
+            "Second text <image> with token.",
+        ]
+        with self.assertRaises(ValueError) as context:
+            processor(text=texts)
+        self.assertTrue("We detected 2 tokens in the text but no images were passed" in str(context.exception))
+
+        # Test with None as Images
+        with self.assertRaises(ValueError) as context:
+            processor(text=text, images=None)
+        self.assertTrue("We detected 1 tokens in the text but no images were passed" in str(context.exception))
+
+        with self.assertRaises(ValueError) as context:
+            processor(text=texts, images=None)
+        self.assertTrue("We detected 2 tokens in the text but no images were passed" in str(context.exception))
+
+    def test_single_tile_image_with_thumbnail_disabled(self):
+        """Test that single-tile images work correctly when use_thumbnail=False."""
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        image_str = "<image>"
+        text_str = "Describe this image."
+        text = image_str + text_str
+
+        # Test with use_thumbnail=False - this should still generate correct tokens
+        inputs = processor(text=text, images=self.small_image, use_thumbnail=False)
+
+        # Count image tokens in input_ids
+        num_image_tokens = sum(1 for token_id in inputs["input_ids"][0] if token_id == self.image_token_id)
+
+        # Verify we have image tokens (the bug caused 0 tokens)
+        self.assertGreater(num_image_tokens, 0, "Single-tile image with use_thumbnail=False should have image tokens")
+
+        # Verify the number of image tokens matches expected based on spatial_shapes
+        spatial_shape = inputs["spatial_shapes"][0].tolist()
+        expected_tokens = math.ceil(spatial_shape[0] / processor.image_processor.downsample_factor) * math.ceil(
+            spatial_shape[1] / processor.image_processor.downsample_factor
+        )
+        self.assertEqual(
+            num_image_tokens,
+            expected_tokens,
+            f"Image tokens ({num_image_tokens}) should match expected ({expected_tokens}) based on spatial shapes",
+        )
+
+        # Verify pixel_values shape is correct
+        encoder_feature_dims = (
+            3 * processor.image_processor.encoder_patch_size * processor.image_processor.encoder_patch_size
+        )
+        self.assertEqual(
+            np.array(inputs["pixel_values"]).shape,
+            (1, processor.image_processor.max_num_patches, encoder_feature_dims),
+        )
+
+    def test_multi_image(self):
+        """Test that text is correctly processed when multiple images are present."""
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        # Text with multiple images and text segments between them
+        text_1 = "First: "
+        text_2 = " Middle: "
+        text_3 = " End."
+        text = text_1 + "<image>" + text_2 + "<image>" + text_3
+        images = [[self.small_image, self.small_image]]
+
+        inputs = processor(text=text, images=images)
+
+        # Construct expected input_ids
+        tokenized_1 = processor.tokenizer(text_1, add_special_tokens=False)["input_ids"]
+        tokenized_2 = processor.tokenizer(text_2, add_special_tokens=False)["input_ids"]
+        tokenized_3 = processor.tokenizer(text_3, add_special_tokens=False)["input_ids"]
+        image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+
+        expected_input_ids = tokenized_1 + image_tokens + tokenized_2 + image_tokens + tokenized_3
+        self.assertEqual(inputs["input_ids"], [expected_input_ids])
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids)])
+
+    def test_multi_turn_multi_image(self):
+        """Test that text is correctly processed when multiple images are present in a multi-turn conversation."""
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding_side="left")
+        processor_components["image_processor"] = self.get_component("image_processor", do_image_splitting=False)
+        processor_kwargs = self.prepare_processor_dict()
+
+        processor = self.processor_class(**processor_components, **processor_kwargs)
+
+        # Simulate a multi-turn conversation with images
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is in image A?"},
+                    {"type": "image"},
+                    {"type": "text", "text": "And image B?"},
+                    {"type": "image"},
+                ],
+            },
+            {
+                "role": "assistant",
+                "content": [{"type": "text", "text": "Image A shows X. Image B shows Y."}],
+            },
+            {
+                "role": "user",
+                "content": [{"type": "text", "text": "Tell me more about image A."}],
+            },
+        ]
+
+        text = processor.apply_chat_template(messages, add_generation_prompt=True)
+        images = [[self.small_image, self.small_image]]
+
+        inputs = processor(text=text, images=images, do_image_splitting=False)
+
+        # Construct expected input_ids based on the chat template structure
+        image_tokens = [self.image_start_token_id] + [self.image_token_id] * 9 + [self.image_end_token_id]
+
+        # Build expected sequence from chat template parts
+        bos = processor.tokenizer(self.bos_token, add_special_tokens=False)["input_ids"]
+        user_start = processor.tokenizer("<|im_start|>user\n", add_special_tokens=False)["input_ids"]
+        assistant_start = processor.tokenizer("<|im_start|>assistant\n", add_special_tokens=False)["input_ids"]
+        im_end = processor.tokenizer("<|im_end|>\n", add_special_tokens=False)["input_ids"]
+
+        text_a = processor.tokenizer("What is in image A?", add_special_tokens=False)["input_ids"]
+        text_b = processor.tokenizer("And image B?", add_special_tokens=False)["input_ids"]
+        assistant_response = processor.tokenizer("Image A shows X. Image B shows Y.", add_special_tokens=False)[
+            "input_ids"
+        ]
+        followup = processor.tokenizer("Tell me more about image A.", add_special_tokens=False)["input_ids"]
+
+        expected_input_ids = (
+            bos
+            + user_start
+            + text_a
+            + image_tokens
+            + text_b
+            + image_tokens
+            + im_end
+            + assistant_start
+            + assistant_response
+            + im_end
+            + user_start
+            + followup
+            + im_end
+            + assistant_start
+        )
+
+        self.assertEqual(inputs["input_ids"], [expected_input_ids])
+        self.assertEqual(inputs["attention_mask"], [[1] * len(expected_input_ids)])