PixVQA — A Unified Video Quality Assessment Toolkit by Pixop

PixVQA is a performance-oriented video quality assessment toolkit for evaluating compression and processing quality across SDR and HDR workflows. It combines established full-reference metrics—PSNR, XPSNR, SSIM, MS-SSIM and VMAF—with PU21-based HDR metrics and optional learned models, including ColorVideoVDP, CompressedVQA-HDR NR, SimpleVQA and HDR-ChipQA. Optimized native CPU processing, optional CUDA acceleration and NVIDIA hardware decoding support efficient evaluation of high-resolution video, with support for both raw YUV and compressed video.

Beyond average quality scores, PixVQA integrates Quality Consistency Score (QCS), a temporal reliability descriptor that uses frame-level or sampled quality scores to quantify how consistently video quality meets defined thresholds. QCS complements conventional metrics by exposing quality drops that an average score can obscure; it is available for metrics that provide suitable per-frame or sampled scores.

Metrics at a Glance

Full-reference metrics need a reference and a distorted video; no-reference metrics score the distorted video alone. The tables below show supported metrics by transfer characteristic. Input requirements (raw YUV, compressed, tags) are stricter than model availability — see Metrics & Formats and Learned Metrics.

Video Format Support by Metric

Metric	Type	SDR – BT.709	PQ / HDR10 – BT.2020	HLG – BT.2020
PSNR	Full-reference	✅	✅	✅
XPSNR	Full-reference	✅	✅	✅
SSIM	Full-reference	✅	✅	✅
MS-SSIM	Full-reference	✅	✅	✅
VMAF (SDR v1.0.16 models)	Full-reference	✅	❌	❌
VMAF (HDR v1.1.8 models)	Full-reference	❌	✅	❌
PU21-PSNR	Full-reference	❌	✅	✅
PU21-SSIM	Full-reference	❌	✅	✅
PU21-MS-SSIM	Full-reference	❌	✅	✅
ColorVideoVDP	Full-reference	✅	✅	✅
CompressedVQA-HDR	No-reference	✅	✅	❌
SimpleVQA	No-reference	✅	❌	❌
HDR-ChipQA	No-reference	✅	✅	🔬

Legend: ✅ Supported · 🔬 Experimental support (validation incomplete) · ❌ Unsupported.

PU21 metrics: PSNR, SSIM and MS-SSIM can be computed after applying the PU21 perceptually uniform transfer function. PU21 transforms HDR luminance values to better approximate the nonlinear sensitivity of human vision. As a result, PU21 provides a perceptually motivated signal representation for conventional full-reference metrics.

Metric Execution Backends

Metric	CPU C/C++	CUDA C++	CPU Python (NumPy/SciPy)	CPU PyTorch	CUDA PyTorch
PSNR	✅	✅	—	—	—
XPSNR	✅	✅	—	—	—
SSIM	✅	✅	—	—	—
MS-SSIM	✅	✅	—	—	—
VMAF	✅	—	—	—	—
PU21-PSNR	✅	✅	—	—	—
PU21-SSIM	✅	✅	—	—	—
PU21-MS-SSIM	✅	✅	—	—	—
ColorVideoVDP	—	—	—	✅	✅
CompressedVQA-HDR	—	—	—	✅	✅
SimpleVQA	—	—	—	✅	✅
HDR-ChipQA	—	—	✅	—	—

✅ Available primary computation backend · — Not implemented for that backend. Python-based metrics may use other libraries for decoding or preprocessing; CUDA C++ refers to native CUDA metric kernels. QCS is a temporal descriptor derived from eligible frame scores, not a separate execution backend.

Install

Supported: Ubuntu 22.04/24.04 x86-64 (including WSL2). Native Windows builds are not validated.

git clone https://github.com/pixop/pixvqa.git
cd pixvqa
bash install_pixvqa.sh
./bin/pixvqa --version

The guided installer lets you choose the features you need:

• Standard installation: Native CPU metrics and software video decoding.
• GPU acceleration (optional): CUDA-accelerated metrics and NVIDIA hardware decoding (NVDEC). Requires CUDA Toolkit 12.8 and a compatible NVIDIA driver.
• AI-based metrics (optional): ColorVideoVDP, CVQA-HDR-NR, SimpleVQA and HDR-ChipQA. These are disabled by default; enable them when prompted, or set PIXVQA_ENABLE_LEARNED=1 for unattended installation.

AI-based metrics use a separate Python 3.12 environment with PyTorch 2.13.0 (CUDA 13.0). HDR-ChipQA also needs a small Python 3.10 environment for its pretrained regression model. You do not need a system-wide CUDA 13 toolkit for the Python metrics, and the installer will not overwrite an existing incompatible Python environment.

See Learned Metrics for detailed requirements and setup.

Use

For compressed video (software-decoder or full build):

./bin/pixvqa --metrics psnr,ssim,msssim \
  -r samples/ref_hevc.mp4 -t samples/dist_hevc.mp4

For planar 10-bit 4:2:0 YUV:

./bin/pixvqa --metrics psnr,ssim --fps 50 \
  -r reference.yuv -t distorted.yuv -s 1920x1080 -f 420p:10

Full-reference comparisons stop at the shorter input and warn if frame counts differ. Use --frames N to cap evaluation explicitly.

Documentation

Installation & Builds · Manual Installation · CLI Usage · Metrics & Formats · Learned Metrics · PU21 · QCS · References · Licensing

Run ./bin/pixvqa --help, --help-advanced or --help-hdr for the authoritative option reference.

License

MIT © 2026 Pixop ApS. See Licensing for third-party and trademark notices.