STARDUSTmark version 3: New forensic watermarking improvements

We’re excited to launch version 3 of STARDUSTmark featuring a practical update to our single-frame forensic watermarking pipeline. The core changes are in embedding and extraction: v3 is more resilient under real distribution transforms, while introducing less visible structure in the image than v2.

This post focuses on measured behavior, not synthetic demos. We cover extraction success/false positives, crop survivability, and visual metrics (PSNR-HVS, SSIM, SSIMULACRA2).

Performance gains

STARDUSTmark v3 consistently outperforms v2, delivering major gains in successful extraction while also reducing false positives across a wide range of content and transformations. At the same time, v3 is typically less visible than v2, with higher PSNR-HVS, SSIM, and SSIMULACRA2 scores across the board. We intentionally don’t use VMAF in this analysis: it was designed and fine-tuned primarily for encoding/compression artifacts, and in our experience it’s not useful for characterizing watermark-induced distortions.

Below are results from our internal comprehensive test suite, which includes a wide variety of content types and distortions (noise, compression artifacts, scaling, cropping). These tests are designed to reflect real-world distribution scenarios, but at significantly higher stress levels.

Cropping robustness is the primary area of improvement in v3. For the results presented below, we progressively cropped each watermarked image and repeated extraction until watermark ID decoding failed, then compared the minimum surviving crop threshold between versions. The reported values represent the average minimum surviving crop.

Here’s an example of how aggressive a crop can be while still allowing watermark ID extraction. This is a best-case scenario based on 3840×1714 content with a 16-bit watermark ID embedded, under ideal conditions (digital capture with very light JPEG compression), so real-world results may vary.

We’ve also increased the maximum capacity of the watermark that can be embedded – v3 can now hold up to 2048 bits of information, compared to the maximum of 256 bits in v2. At the same time, extraction speed improved as well, with v3 running about 3.1 times faster on average than v2 across the same internal comprehensive test suite referenced above.

New HVS-based perception model

STARDUSTmark v3 introduces a new human visual system (HVS) based perception model that guides where and how watermark energy is embedded. It analyzes local image structure, texture, and masking characteristics to find embedding areas that can carry stronger signal without becoming perceptible, maximizing robustness while minimizing visible impact. The new model keeps the watermark effectively invisible even under scrutiny, while removing the subtle patterning that could occasionally appear in flat, banding-prone regions with STARDUSTmark v2.

For a comparison, we used STARDUSTmark v3, STARDUSTmark v2, and a top-tier open-source ML watermarking solution released in late 2025 on the ASC Standard Evaluation Material 2 (StEM2) video (Rec.709, 1920×1080 rendition).

The table summarizes the aggregate fidelity metrics.

See the improvements for yourself

Across the tests shown here, v3 improves the reliability/visibility tradeoff in a meaningful way: higher extraction robustness, lower false positives, and cleaner output under visual inspection. In short, v3 gives us more watermark headroom where it matters, with less artifact risk where viewers actually notice it.

Reach out to us if you’d like to see focused side-by-side examples to show what the numbers in this post look like in actual frames.