diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml
index 0a718ba6d512..0edb00aacd6a 100644
--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -26,6 +26,8 @@
       title: Exporting to production
     - local: modular_transformers
       title: Contributing a new model to Transformers
+    - local: add_vision_processing_components
+      title: Add vision processing components
     - local: add_new_model
       title: Legacy model contribution
     - local: auto_docstring
diff --git a/docs/source/en/add_vision_processing_components.md b/docs/source/en/add_vision_processing_components.md
new file mode 100644
index 000000000000..a77a19a36d1d
--- /dev/null
+++ b/docs/source/en/add_vision_processing_components.md
@@ -0,0 +1,88 @@
+<!--Copyright 2026 The HuggingFace 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Add vision processing components
+
+Adding a vision language model (VLM) requires two image processor classes on top of the standard [modular](./modular_transformers) approach.
+
+> [!NOTE]
+> For the modeling and config steps, follow the [modular](./modular_transformers) guide first.
+
+- [torchvision](https://docs.pytorch.org/vision/stable/index.html) backend is the default and supports GPU acceleration.
+- [PIL](https://pillow.readthedocs.io/en/stable/index.html) backend is a fallback when no GPU is available.
+
+Both classes share the same preprocessing logic but have different backends. Their constructor signatures and default values must be identical. [`AutoImageProcessor.from_pretrained()`] selects the backend at load time and falls back to PIL when torchvision isn't available. Mismatched signatures cause the same saved config to behave differently across environments.
+
+## torchvision
+
+Create `image_processing_<model_name>.py` with a class that inherits from [`TorchvisionBackend`]. Define a kwargs class first if your processor needs custom parameters beyond the standard [`ImagesKwargs`].
+
+```py
+from ...image_processing_backends import TorchvisionBackend
+from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import auto_docstring
+
+class MyModelImageProcessorKwargs(ImagesKwargs, total=False):
+    tile_size: int  # any model-specific kwargs
+
+@auto_docstring
+class MyModelImageProcessor(TorchvisionBackend):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"shortest_edge": 224}
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+
+    def __init__(self, **kwargs: Unpack[MyModelImageProcessorKwargs]):
+        super().__init__(**kwargs)
+```
+
+## PIL
+
+Create `image_processing_pil_<model_name>.py` with a class that inherits from [`PilBackend`]. Import the kwargs class from the torchvision file, but don't redefine it. Sharing the same class keeps both backends' kwargs in sync. For processors with no custom parameters, use [`ImagesKwargs`] directly.
+
+```py
+from ...image_processing_backends import PilBackend
+from ...image_utils import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD, PILImageResampling
+from ...utils import auto_docstring
+from .image_processing_<model_name> import MyModelImageProcessorKwargs
+
+@auto_docstring
+class MyModelImageProcessorPil(PilBackend):
+    resample = PILImageResampling.BICUBIC
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"shortest_edge": 224}
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = True
+
+    def __init__(self, **kwargs: Unpack[MyModelImageProcessorKwargs]):
+        super().__init__(**kwargs)
+```
+
+> [!TIP]
+> See [`CLIPImageProcessor`]/[`CLIPImageProcessorPil`] and [`LlavaOnevisionImageProcessor`]/[`LlavaOnevisionImageProcessorPil`] for reference.
+
+## Next steps
+
+- Read the [Auto-generating docstrings](./auto_docstring) guide to auto-generate consistent docstrings with `@auto_docstring`.
+- Read the [Writing model tests](./testing) guide to write integration tests for your model.