python-nvidia-codec

A Pythonic library for NVIDIA Video Codec (NVDEC).

The project is still in active development; expect breaking changes.

Requirements

• Python 3.10+
• NVIDIA GPU with NVDEC support
• NVIDIA driver with libnvcuvid.so
• FFmpeg 8.x (for demuxing)

Installation

pip install nvidia-codec

Or from source:

git clone https://github.com/riaqn/python-nvidia-codec.git
cd python-nvidia-codec
pip install -e .

Usage

Screenshot (single frame extraction)

from nvidia_codec.utils import Screenshoter
from datetime import timedelta
import torch

# Optional: resize frames to fit within target dimensions
def target_size(h, w):
    scale = min(1920 / w, 1080 / h, 1.0)
    return (int(h * scale) // 2 * 2, int(w * scale) // 2 * 2)

player = Screenshoter('/path/to/video.mp4', target_size=target_size, device=0)
print(f'Duration: {player.duration}')

# Get frame at 10 seconds (fast keyframe seek by default)
timestamp, frame = player.screenshot(timedelta(seconds=10), torch.uint8)

# For precise frame timing, use accurate=True (slower)
timestamp, frame = player.screenshot(timedelta(seconds=10), torch.uint8, accurate=True)

player.free()

Decode all frames

from nvidia_codec.utils import Player
import torch

player = Player('/path/to/video.mp4', device=0)

for timestamp, frame in player.frames(torch.uint8):
    # frame is a torch tensor on GPU (CHW format)
    print(f'{timestamp}: {frame.shape}')

Supported Codecs

Codec support depends on your GPU:

• H.264 (AVC)
• H.265 (HEVC)
• VP9
• AV1 (Ampere+)
• VC1/WMV3 (older GPUs only, removed in Turing)

Why another Python library for NVIDIA Codec?

Comparison to Video-Processing-Framework (VPF)

Methodologies: VPF is written fully in C++ and uses pybind to expose Python interfaces. PNC is written fully in Python and uses ctypes to access NVIDIA C interfaces. Our code tends to be more concise, less duplicative and easier to read and write. It also allows better interoperability with other Python libraries.

Performance: Preliminary tests show little to no difference in performance, because the heavy lifting is done on the GPU anyway. Both libraries can saturate the GPU decoder. PNC uses more CPU than VPF as expected from Python vs. C++, but still negligible (less than 10% of Ryzen 3100 single core for 8K×4K HEVC).

Resource Management:

• In VPF, Surface given to user is not owned by the user. It will be overwritten by new frames which is counter-intuitive. Picture is not exposed to user at all - they are always mapped (post-processed and copied) to Surface so the picture can be ready for new frames. The latter is inefficient when only a subset of Pictures are needed (e.g., screenshots).
• VPF allocates the bare minimum of resources needed for most decoding tasks. PNC allows the user to specify the amount of resources to be allocated for advanced applications. Users own the resources and decide when and whether to deal with them.
• Managing resources is not painful: similar to pycuda, we shift the burden of managing host/device resources to the Python garbage collector. Resources (such as Picture and Surface) are automatically freed when the user drops the reference.

Roadmap

• Decoding
• Color Conversion
  • Source Format: NV12, P016, YUV444, YUV444_16Bit
  • Target Format: RGB24, RGB48, RGB444P, RGB444P16
  • Color Ranges: MPEG (limited), JPEG (full)
  • Color Space: BT.601, BT.709, BT.2020
• Built-in cropping and scaling via NV decoder
• Thread-safe and thread-friendly
• Encoder
• PyTorch integration

Acknowledgements

• Many thanks to @rarzumanyan for all the help and explanations!
• The blog posts from myMusing are very helpful.