Looking to implement C2PA? Trufo provides tooling to take care of everything from certificates and timestamping to watermarking and fingerprinting. Learn More

A New Gold Standard for Watermarks

Introducing PawPrint

The first neural watermark designed for provenance at scale.

Published July 22, 2026

We are introducing our new line of AI-powered watermarks, PawPrint, available for image and audio and video. PawPrint represents a massive leap in performance over existing models, across all dimensions (speed, accuracy, durability, perceptibility, etc.). It can be swiftly deployed in any business use case where digital media content needs to be reliably identified. PawPrint is available today as part of our C2PA Signing API & SDK services.

Overview

Trufo's pawprint.image watermark delivers a substantial improvement in watermark performance across the board:

Watermark	Company	Year	Payload [bits]	Decode [ms]	LPIPS¹ ↓	FPR ↓	Durability ↑	Score
pawprint.image	Trufo	2026	44A-	1.93A	0.0027A-	0.000A	0.725A-	94
pixelseal²	Meta	2025	128A	4.38A-	0.0022A-	0.001B+	0.645B	86
vine²	Academic	2025	64A-	7.14B+	0.0148C+	0.003B+	0.488C+	70
trustmark.p	Adobe	2024	61A-	3.69A-	0.0017A-	0.001B+	0.441C	78
trustmark.q	Adobe	2024	61A-	4.01A-	0.0020B	0.001B+	0.528C+	76
jpeg2026	Trufo³	2024	100A	24.91B-	0.0017A	0.006B	0.611B	80
dctdwtsvd	Traditional	2021	64A-	45.32C+	0.0206C	0.954D	0.248D+	42

¹Assigned grades are adjusted based on a mix of other perceptibility metrics.

²Enhanced by Trufo with BCH error correction.

³This is a traditional watermark (pawprint v2024b), and was a winner of the 2026 JPEG Trust Grand Challenge.

When paired with Trufo's provenance platform, pawprint.image becomes the natural choice for practical watermark deployment in almost any commercial setting.

The Watermark

In 2024, Trufo introduced the original pawprint watermark (the JPEG-2026 version with a 100-bit payload is recorded above). This was a breakthrough traditional watermark, completely surpassing the watermarks that came before.

Since then, neural network capabilities have progressed substantially. As of 2026, two neural watermarks now outperform the original pawprint in a holistic setting, one of them being Trufo's new pawprint.image neural watermark. Compared to the original pawprint, the new pawprint.image is faster to run (GPUs) and stronger on flat digital styles.

Here is an example of the pawprint.image watermark applied on a typical image.

Evaluation

We compare pawprint.image against six popular watermarks that we were able to obtain for testing:

pixelseal (Meta, 2025)
vine (Academic, 2025)
trustmark.p (Adobe, 2024)
trustmark.q (Adobe, 2024)
jpeg2026 (Trufo, 2024)
dctdwtsvd (Traditional, 2021)

The evaluations are conducted on a fixed 1,000-image mix of drawn from five datasets: domain.net, stylebreeder, unsplash, mitsua, and the JPEG-2026 set. The test set was excluded from the training set.

Speed

We benchmarked encode and decode speed at native image resolution at steady state after warm-up.

	pawprint.image	pixelseal	vine	trustmark.p	trustmark.q	jpeg2026	dctdwtsvd
encode (ms)	10.12	6.84	51.82	8.13	8.81	372.46	69.67
decode (ms)	1.93	4.38	7.14	3.69	4.01	24.91	45.32

The pawprint.image decoder is the fastest; the bottleneck is reading the image into a tensor. This means that, for systems where the decoder needs to check multiple watermarks, pawprint.image is the least costly to run.

Perceptibility

We used four metrics: PSNR, SSIM, LPIPS, and wM4E.

	pawprint.image	pixelseal	vine	trustmark.p	trustmark.q	jpeg2026	dctdwtsvd
PSNR [dB] ↑	43.86	46.19	35.10	48.06	42.80	48.92	40.41
SSIM ↑	0.99360	0.99410	0.98632	0.99612	0.99155	0.99659	0.98433
LPIPS ↓	0.00265	0.00219	0.01484	0.00174	0.00198	0.00175	0.02062
wM4E ↓	0.274	0.465	164,089	1.235	16.375	0.088	12.356

Note that these metrics do not fully capture perceptibility. Training a neural network discriminator is typically the strongest measure of perceptibility.

The wM4E formula is:

\mathrm{wM4E}(R,W) = \Big\langle \sum_{c \in \{Y,Cr,Cb\}} \alpha_c\!\left(\frac{C_c(W)-C_c(R)}{\sigma(R)}\right)^4 \Big\rangle_{x,y}, \qquad (\alpha_Y,\alpha_{Cr},\alpha_{Cb}) = (1,0.5,0.5)

Here, σ(R) is a measure of local perceptual masking capacity, and α are luma/chroma sensitivity coefficients.

Durability

We tested exact recovery of the payload against a combination of attacks. The evaluation included the following attacks:

editing: overlay / occlusion
filters: blur, grayscale, various color transformations
geometry: rotation (orthogonal and small-angle), reflection, resizing, cropping
compression: JPEG, WebP, JPEG 2000 (wide range of Q)

Unlike benchmarks that apply one easy attack at a time, these attacks fire independently, so a test image can be subjected to any number of attacks. This provides a more practical evaluation of watermark durability. We focus on exact payload identification: after all, if returned payload is incorrect, the watermark cannot be used to retrieve provenance data.

Breaking down the durability evaluation:

by attack count

0 attacks

1

2

3+

by single attack

	pawprint.image	pixelseal	vine	trustmark.p	trustmark.q	jpeg2026	dctdwtsvd
Overlay	0.697	0.743	1.000	0.257	0.286	0.811	0.491
Blur	0.962	0.833	0.897	0.987	1.000	0.577	0.359
Color	0.959	0.954	0.981	0.874	0.939	0.937	0.276
Grayscale	1.000	1.000	1.000	0.253	0.928	0.988	0.000
Rotate & Reflect	1.000	0.726	0.000	0.156	0.154	0.973	0.057
Resize	1.000	1.000	1.000	0.988	1.000	0.951	0.000
Rotate Angle	0.786	0.801	0.015	0.180	0.272	0.646	0.029
Side crop	1.000	0.995	0.033	0.973	0.984	0.802	0.121
Axis crop	0.518	0.507	0.000	0.157	0.212	0.288	0.000
JPEG	0.663	0.606	0.943	0.460	0.782	0.608	0.248
WebP	0.484	0.333	0.968	0.452	0.731	0.333	0.097
JPEG 2000	0.570	0.329	0.797	0.582	0.671	0.316	0.354

Trufo's pawprint.image has the highest overall durability in this evaluation.

Overall Score

The overall ranking of tested watermarks, based on our composite score, is as follows:

pawprint.image

94

pixelseal

86

jpeg2026

80

competitor est.

80

trustmark.p

78

trustmark.q

76

vine

70

legacy est.

50

dctdwtsvd

42

There are a number of commercial watermarks we were unable to test. We estimate that the strongest competitors would score around 80, at the upper end of academia (only Trufo and Meta have exceeded the 80 threshold). Legacy watermark systems, from companies using older technology or lacking AI research teams, would likely score around 50.

Overall, even if accounting for different methods of evaluation, the clear winner is pawprint.image.