Configuring TrustMark

Overview

All watermarking algorithms trade off between three properties:

• Capacity (bits)
• Robustness (to various transformations)
• Visibility (of watermark)

This document explains how to configure TrustMark to tune these properties, however the default configuration for TrustMark (variant Q, 100% strength, BCH_5 error correction) is sufficient for most use cases.

Model variant

TrustMark has four model variants (B, C, P, and Q) that may be selected when instantiating TrustMark. All encode/decode calls on the object will use this variant.

In general, we recommend using P or Q:

• P is useful for creative applications where very high visual quality is required.
• Q is a good all-rounder and is the default.

Note: Images encoded with one model variant cannot be decoded with another.

Variant	Typical PSNR	Model Size Enc/Dec (MB)	Description
Q	43-45	17/45	Default (Quality). Good trade-off between robustness and imperceptibility. Uses ResNet-50 decoder.
B	43-45	17/45	(Beta). Very similar to Q, included mainly for reproducing the paper. Uses ResNet-50 decoder.
C	38-39	17/21	(Compact). Uses a ResNet-18 decoder (smaller model size). Slightly lower visual quality.
P	48-50	16/45	(Perceptual). Very high visual quality and good robustness. ResNet-50 decoder trained with much higher weight on perceptual loss (see paper).

Watermark strength

Set the optional WM_STRENGTH parameter when encoding (at runtime). Its default value is 1.0, and changing it provides a trade-off between robustness and visibility:

• Raising its value (for example, to 1.5) improves robustness (so, for example, the watermark survives printing) but increases the likelihood of ripple artifacts.
• Lowering its value (for example, to 0.8) reduces any likelihood of artifacts but compromises on robustness; however it still survives lower noise, screenshotting, or social media.

For example:

encoded_image = tm.encode(input_image, payload="example", WM_STRENGTH=1.5)

Error correction level

TrustMark encodes a payload (the watermark data embedded within the image) of 100 bits. The data schema implemented in python/datalayer.py enables you to choose an error correction level over the raw 100 bits of payload to maintain reliability under transformations or noise.

Encoding modes

Set the error correction level using one of the four encoding modes.

The following table describes TrustMark's encoding modes:

Encoding	Protected payload	Number of bit flips allowed
Encoding.BCH_5	61 bits (+ 35 ECC bits)	5
Encoding.BCH_4	68 bits (+ 28 ECC bits)	4
Encoding.BCH_3	75 bits (+ 21 ECC bits)	3
Encoding.BCH_SUPER	40 bits (+ 56 ECC bits)	8

Specify the mode when you instantiate the encoder, as follows:

tm=TrustMark(verbose=True, model_type='Q', encoding_type=TrustMark.Encoding.<ENCODING>)

Where <ENCODING> is BCH_5, BCH_4, BCH_3, or BCH_SUPER.

For example:

tm=TrustMark(verbose=True, model_type='Q', encoding_type=TrustMark.Encoding.BCH_5)

The decoder automatically detects the data schema in a watermark, so you can choose the level of robustness that best suits your use case.

Selecting the model and strength implicitly selects the level of robustness and visibility of the watermark. If you have reduced robustness for lower visibility, you can regain some robustness by increasing error correction (at the cost of payload capacity). Note that even 40 bits gives a key space of around one trillion.

Center cropping

TrustMark generates residuals at 256 x 256 and then scales/blends them into the original image. Several derivative papers have adopted this universal resolution-scaling trick.

• If the original image is extremely long/thin (aspect ratio beyond 2:1), the residual watermark will degrade when scaled.
• TrustMark addresses this by automatically center-cropping the image to a square if the aspect ratio exceeds 2.0. For example, for a 1000 x 200 image, only a 200 x 200 region in the center carries the watermark.
• The aspect ratio limit can be overridden via the ASPECT_RATIO_LIM parameter. Setting it to 1.0 always forces center-crop behavior (useful for content platforms that square-crop images). This is the default when using model variant P.

Do not concentrate watermarks

Visual quality is often measured via PSNR (Peak Signal-to-Noise Ratio), but PSNR does not perfectly correlate with human perception of watermark visibility.

Some derivative works have tried to improve PSNR by "zero padding" or concentrating the watermark into only the central region, effectively altering fewer pixels and artificially raising the PSNR score. However, this approach can increase human-visible artifacts in the concentrated area.

Parameter: CONCENTRATE_WM_REGION

• Default is 100% (no zero padding).
• If you set, for example, 80% zero padding, you might inflate PSNR by ~5 dB.
• At 50% zero padding, PSNR might inflate by ~10 dB.
• In some extreme cases, PSNR can reach 55–60 dB but at the cost of noticeable artifacts in that smaller region.

In summary: While the functionality exists for comparison purposes, it’s not recommended for production. High concentration setting yields high PSNR but paradoxically more visible artifacts in the watermarked area.