nvdec

GPU-resident runtime for TensorRT image-to-image video models

GPU-accelerated video I/O nodes for ComfyUI, backed by Nelux (NVDEC decode / NVENC encode). ~11x faster 1080p transcode, 13x less RAM.

High-performance Python architecture for multi-stream NVDEC decoding and GPU inference using DLPack and PyTorch CUDA IPC to bypass the GIL.

GPU-accelerated RIFE AI video frame interpolation app that converts 24/30 FPS footage to smooth 60/120/144 FPS with CUDA, TensorRT and NVENC.

GPU-accelerated video pipeline — honest benchmarks (OpenCV CUDA/NVDEC vs GStreamer), YOLO11 + ByteTrack analytics, multi-camera batching, FP16/TensorRT, Prometheus

Swift port of an NVIDIA DeepStream object detector running on a Jetson Orin Nano. ~Half the CPU of the Python version on the same DeepStream graph.

Drop-in cv2.VideoCapture replacement with a selectable decode backend (CPU, CUDA/NVDEC, OpenCV, or DeepStream) for RTSP and file streams.

High-fidelity tensor alignment framework between CPU-based OpenCV decoding and GPU-accelerated NVDEC (VALI) pipelines. Eliminates inference divergence and tracking instability in production-grade YOLO deployments.

Unmanic plugin: H.265/HEVC encoder using NVIDIA hevc_nvenc with a true end-to-end GPU pipeline. Fork of Josh5/unmanic.plugin.encoder_video_hevc_nvenc that adds -hwaccel_output_format cuda when NVDEC HW decoding is enabled, keeping decoded frames in GPU memory through NVENC. Drop-in replacement with sensible defaults and full settings parity.