Research Paper · March 2026v1.2

Real-World Watermark Benchmark

We ran 476 decode tests through real WordPress, Cloudinary CDN, and a WAM architecture benchmark — to find which watermark survives actual web infrastructure and whether the 150px limit can be overcome.

Full methodology, raw results, and the data-driven decision behind our production default. Every number published. Every limitation documented.

476
Decode tests executed
91%
Recovery rate (token40 Q)
98%
WAM recovery rate (all sizes)
3
Watermark architectures compared
Background

Why this research matters

Lab tests aren't enough

Most watermark benchmarks test JPEG compression in isolation. Real web infrastructure applies unpredictable chains of resize, re-encode, and format conversion.

! WordPress serves 43% of the web

If your watermark doesn't survive WordPress's image pipeline, it fails for nearly half the internet. We tested on a real WordPress instance.

Transparent methodology

Every test parameter, every limitation, and every result is documented. We publish what failed, not just what worked.

Contents

What's in the paper

01
Executive Summary
v1.1: 150×150 limit confirmed · v1.2: WAM benchmark — 150px limit overcome (10/11, 91%)
02
Methodology
476 total tests: 366 real-world + 55 TrustMark thumbnail + 55 WAM benchmark
03
WordPress Rendition Results
Recovery rates by image size — 300px floor for TrustMark Q, 150px achieved by WAM
04
Cloudinary CDN Transform Results
Aggressive transforms: resize, crop, quality reduction, format conversion
05
Per-Image Performance
Individual image breakdown across all candidates
06
Decision & Implementation
Why token40_Q is current default — WAM identified as v2 embedding candidate
07
Limitations & Reproducibility
wm_strength, tiling, WAM — all three follow-up hypotheses concluded
Results

Key findings

token40_Q achieves 91% recovery across real infrastructure

111 of 122 WordPress and Cloudinary renditions were correctly decoded — a +2% improvement over the previous baseline with zero regressions on any image or transform.

WAM overcomes the 150px limit — 10/11 (91%) where TrustMark achieves 7/11 (64%) (v1.2)

WAM (Watermark Anything, Meta AI / ICLR 2025) was benchmarked on the same 11-image × 5-size matrix. All 4 TrustMark-specific 150px failures were resolved. WAM is identified as the clear candidate for a next-generation embedding layer.

150×150 TrustMark limit is physical — increasing wm_strength yields 0/4 improvements (v1.1)

A controlled software-only benchmark and a wm_strength=1.5 spot-test confirmed the limit is architectural, not configuration. The 300px floor (WordPress Medium) remains the TrustMark recovery boundary.

Q and P variants serve different use cases

Q (frequency-domain DCT perturbations) excels at compression resilience — ideal for web publishing. P (redundant spatial blocks) excels at crop resilience — ideal for social media and e-commerce.

Full analysis: We Put Our Watermarks Through Real WordPress and CDN Pipelines →

Engineering rigor you can verify

Read the full paper, check our methodology, then try MarkMyAI yourself.

Questions about methodology? hello@markmyai.com
Analytics Consent

We use Google Analytics 4 only if you agree, to understand which pages bring traffic and where visitors drop off. No advertising features are enabled. You can change your choice at any time in the privacy settings.