Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/pytorch/vision/llms.txt

Use this file to discover all available pages before exploring further.

TorchVision is PyTorch’s official computer vision library, maintained by the PyTorch team and contributors since 2016. It bundles the most commonly needed building blocks for computer vision workflows — from loading raw images and videos to applying state-of-the-art pre-trained models — all deeply integrated with PyTorch’s tensor ecosystem and autograd engine. Whether you are training a custom model from scratch, fine-tuning a pre-trained backbone, or running inference in production, TorchVision provides the primitives to get the job done without reinventing the wheel.

​
What’s Inside TorchVision

TorchVision is organized into six top-level modules, each covering a distinct concern in a typical computer vision pipeline.

torchvision.models

Pre-trained architectures for classification, detection, segmentation, optical flow, and video understanding. Weights are versioned via the WeightsEnum API introduced in v0.13.

torchvision.transforms

Image and video augmentation pipelines. The modern v2 API operates on PIL images, tensors, and the new TVTensors types (masks, bounding boxes, keypoints) in a single, composable pass.

torchvision.datasets

Ready-to-use torch.utils.data.Dataset implementations for dozens of popular benchmarks including ImageNet, COCO, VOC, CIFAR, and more.

torchvision.ops

Computer-vision–specific operators such as nms, roi_align, deform_conv2d, and box_iou that are not present in core PyTorch.

torchvision.io

Low-level image and video reading and writing backed by native C++ codecs. Supports JPEG, PNG, WebP, and video via PyAV (FFmpeg bindings).

torchvision.utils

Visualization helpers such as make_grid, draw_bounding_boxes, draw_segmentation_masks, and draw_keypoints for inspecting model outputs.

​
Modern Transforms and TVTensors

TorchVision v0.15 introduced a redesigned torchvision.transforms.v2 API alongside a new type system called TVTensors. TVTensors are thin torch.Tensor subclasses — Image, BoundingBoxes, Mask, and Video — that carry semantic metadata so that a single transform pipeline can correctly handle every tensor type simultaneously. For example, a RandomHorizontalFlip applied to a (image, BoundingBoxes, Mask) tuple will flip all three in a geometrically consistent way. 
import torchvision.transforms.v2 as T
from torchvision.tv_tensors import BoundingBoxes, Mask

transform = T.Compose([
    T.RandomHorizontalFlip(p=0.5),
    T.RandomCrop(size=(224, 224)),
    T.ToDtype(torch.float32, scale=True),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
The v2 transforms API is the recommended path for all new projects. The legacy torchvision.transforms (v1) namespace remains available for backwards compatibility but will not receive new features.

​
Version Compatibility

TorchVision releases are tightly coupled to PyTorch releases. Always install matching versions using the official PyTorch installer at pytorch.org/get-started/locally. The table below shows recent stable pairings and their supported Python ranges.
torchtorchvisionPython
main / nightlymain / nightly>=3.10, <=3.14
2.120.27>=3.10, <=3.14
2.110.26>=3.10, <=3.14
2.100.25>=3.10, <=3.14
2.90.24>=3.10, <=3.14
2.80.23>=3.9, <=3.13
2.70.22>=3.9, <=3.13
2.60.21>=3.9, <=3.12
Mismatched torch and torchvision versions are one of the most common sources of runtime errors. Always verify with import torchvision; print(torchvision.__version__) after installation.

​
Image Backends

TorchVision supports multiple image-loading backends, selectable at runtime via torchvision.set_image_backend():
  • PIL / Pillow (default) — The reference backend. Supports the widest range of image formats and operations.
  • Pillow-SIMD — A drop-in replacement for Pillow that uses SIMD CPU instructions for significantly faster decoding. Swap in without any code changes.
  • accimage — Uses the Intel IPP library. Generally faster than Pillow for JPEG decoding but supports fewer operations and formats.
import torchvision

# Switch to accimage for faster JPEG loading (requires accimage to be installed)
torchvision.set_image_backend("accimage")

# Query the active backend at any time
print(torchvision.get_image_backend())  # "accimage"
For most workflows the default PIL backend is sufficient. Consider Pillow-SIMD or accimage only when image loading is a measured bottleneck in your training pipeline.

​
License

TorchVision is released under the BSD 3-Clause License. Pre-trained model weights may carry additional terms derived from the datasets on which they were trained — for example, SWAG models are released under CC-BY-NC 4.0. Always verify the license of any specific weights before commercial use.

Build docs developers (and LLMs) love

Get started for freeTalk to us