approxSSIMate

Lightweight models for approximating SSIM from global distortion signals.

Presented at QoMEX 2026: poster PDF

approxSSIMate provides fast, reference-based models that approximate SSIM using only:

• Global MSE (or PSNR)
• Local statistics of the reference image

The goal is to enable SSIM-like reasoning without computing full window-based SSIM over both images.

News

• June 2026: Initial research-preview release, v0.1.0 “Cardiff”, prepared for QoMEX 2026.

Installation

Install the current research-preview version from source:

git clone https://github.com/luctrudeau/approxSSIMate.git
cd approxSSIMate
pip install -e .

Quick start

approxSSIMate uses a two-step workflow:

1. Compute a source-dependent calibration file (.k) from the reference image

approxssimate k MY_IMAGE.png -o MY_IMAGE.k

2. Use the .k to estimate SSIM from MSE values or distorted images. You can specify either the MSE directly

approxssimate ssim -k MY_IMAGE.k --mse="12.58,28.17,41.52,53.66" 

Or instead you can specify the distorted image(s)

approxssimate ssim -k MY_IMAGE.k MY_IMAGE.png MY_DISTORTED_IMG_1.png MY_DISTORTED_IMG_2.png MY_DISTORTED_IMG_3.png

Why?

Computing SSIM requires local window statistics from both the reference and distorted image.

In many practical encoding scenarios:

• The reference image is fixed
• Multiple distorted versions are evaluated
• Only global distortion (MSE / PSNR) is available

Examples include:

• Bitrate ladder construction
• Multi-encoding experiments
• Convex-hull selection workflows
• Fast rate–distortion exploration

In such cases, recomputing full SSIM repeatedly can be expensive.

approxSSIMate provides fast approximations that reuse reference-image statistics and operate from global MSE only.

Notes

• Images are converted to grayscale internally.
• All distorted images must match the reference resolution.
• Designed for batch evaluation workflows (e.g., bitrate ladder construction).

Sponsorship

approxSSIMate is an open research project focused on making perceptual quality evaluation faster and more practical for real-world encoding workflows.

If your organization benefits from faster SSIM estimation, large-scale encoding experiments, bitrate ladder construction, or convex-hull optimization workflows, consider sponsoring the project.

Roadmap (Possible sponsored milestones)

Tier 1 — Native C Implementation

Develop a production-ready C implementation of the SSIM approximation models:

• Optimized for speed and low memory footprint
• Designed for production integration
• Architecture compatible with libVMAF’s SSIM implementation
• Potential upstream contribution to libVMAF
• Also available as a standalone CLI and embeddable library

Tier 2 — SIMD Optimizations

Architecture-specific acceleration layers:

• AVX2 / AVX-512 (x86 servers)
• NEON (ARM-based systems)

Support & Collaboration

• Sponsor via GitHub Sponsors
• Reach out directly to discuss collaboration or production integration