Watermarking Algorithms

MarkDiffusion supports 11 state-of-the-art watermarking algorithms for latent diffusion models.

Algorithm Overview

Algorithm	Category	Target	Description
Tree-Ring (TR)	Pattern	Image	Embeds invisible ring patterns in frequency domain
Ring-ID (RI)	Pattern	Image	Multi-key identification with tree-ring patterns
ROBIN	Pattern	Image	Robust and invisible watermarks with adversarial optimization
WIND	Pattern	Image	Two-stage robust watermarking hidden in noise
SFW	Pattern	Image	Semantic watermarking with Fourier integrity
Gaussian-Shading (GS)	Key	Image	Provable performance-lossless image watermarking
GaussMarker (GM)	Key	Image	Robust dual-domain watermarking
PRC	Key	Image	Undetectable watermark for generative models
SEAL	Key	Image	Semantic-aware image watermarking
VideoShield	Key	Video	Video diffusion model regulation via watermarking
VideoMark	Key	Video	Distortion-free robust watermarking for video

Pattern-Based Methods

Pattern-based methods embed predefined patterns into the generation process.

Tree-Ring (TR)

Reference: Tree-Ring Watermarks: Fingerprints for Diffusion Images that are Invisible and Robust

Tree-Ring embeds circular patterns in the Fourier domain of initial latents, making them invisible in the spatial domain but detectable through frequency analysis.

Key Features:

• Invisible watermarks in spatial domain

• Robust to common image transformations

• No need for additional neural networks

Usage:

from markdiffusion.watermark.auto_watermark import AutoWatermark

watermark = AutoWatermark.load(
    'TR',
    algorithm_config='markdiffusion/config/TR.json',
    diffusion_config=diffusion_config
)

# Generate watermarked image
image = watermark.generate_watermarked_media(prompt)

# Detect watermark
result = watermark.detect_watermark_in_media(image)

Configuration Parameters:

From markdiffusion/config/TR.json:

• w_seed: 999999 - Watermark seed

• w_channel: 0 - Channel index to embed watermark

• w_pattern: “zeros” - Pattern type

• w_mask_shape: “circle” - Mask shape

• w_radius: 10 - Ring radius in frequency domain

• w_pattern_const: 0 - Pattern constant

• threshold: 50 - Detection threshold

Ring-ID (RI)

Reference: RingID: Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification

Ring-ID extends Tree-Ring to support multiple keys, enabling multi-user identification and improved robustness.

Key Features:

• Multi-key identification

• Enhanced robustness

• Backward compatible with Tree-Ring

Usage:

watermark = AutoWatermark.load(
    'RI',
    algorithm_config='markdiffusion/config/RI.json',
    diffusion_config=diffusion_config
)

Configuration Parameters:

From markdiffusion/config/RI.json:

• ring_width: 1 - Ring width

• quantization_levels: 4 - Quantization levels

• ring_value_range: 64 - Ring value range

• assigned_keys: 10 - Number of assigned keys

• radius: 14 - Ring radius

• radius_cutoff: 3 - Radius cutoff

• heter_watermark_channel: [0] - Heterogeneous watermark channel

• ring_watermark_channel: [3] - Ring watermark channel

• threshold: 50 - Detection threshold

ROBIN

Reference: ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization

ROBIN uses adversarial optimization to create robust watermarks that are invisible to human eyes and resistant to attacks.

Key Features:

• Adversarial optimization for robustness

• Invisible watermarks

• Trained watermark generator

Usage:

watermark = AutoWatermark.load(
    'ROBIN',
    algorithm_config='markdiffusion/config/ROBIN.json',
    diffusion_config=diffusion_config
)

Configuration Parameters:

From markdiffusion/config/ROBIN.json:

• w_seed: 999999 - Watermark seed

• w_channel: 3 - Watermark channel

• w_pattern: “ring” - Pattern type

• w_up_radius: 30 - Upper radius

• w_low_radius: 5 - Lower radius

• watermarking_step: 35 - Watermarking injection step

• threshold: 45 - Detection threshold

• learning_rate: 0.0005 - Training learning rate

• max_train_steps: 2000 - Maximum training steps

WIND

Reference: Hidden in the Noise: Two-Stage Robust Watermarking for Images

WIND implements a two-stage watermarking approach that hides watermarks in the noise initialization.

Key Features:

• Two-stage watermarking

• Hidden in initial noise

• High robustness

Usage:

watermark = AutoWatermark.load(
    'WIND',
    algorithm_config='markdiffusion/config/WIND.json',
    diffusion_config=diffusion_config
)

SFW

Reference: Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity

SFW combines semantic information with Fourier domain watermarking for enhanced robustness and quality.

Key Features:

• Semantic-aware watermarking

• Fourier integrity preservation

• Minimal quality degradation

Usage:

watermark = AutoWatermark.load(
    'SFW',
    algorithm_config='markdiffusion/config/SFW.json',
    diffusion_config=diffusion_config
)

Key-Based Methods

Key-based methods use secret keys to embed and extract watermarks.

Gaussian-Shading (GS)

Reference: Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models

Gaussian-Shading provides provably performance-lossless watermarking by injecting Gaussian noise into the generation process.

Key Features:

• Performance-lossless (provable)

• No quality degradation

• Simple and efficient

Usage:

watermark = AutoWatermark.load(
    'GS',
    algorithm_config='markdiffusion/config/GS.json',
    diffusion_config=diffusion_config
)

Configuration Parameters:

From markdiffusion/config/GS.json:

• channel_copy: 1 - Channel to copy watermark

• wm_key: 42 - Watermark key

• hw_copy: 8 - Height/width copy parameter

• chacha: true - Use ChaCha encryption

• chacha_key_seed: 123456 - ChaCha key seed

• chacha_nonce_seed: 789012 - ChaCha nonce seed

• threshold: 0.7 - Detection threshold

GaussMarker (GM)

Reference: GaussMarker: Robust Dual-Domain Watermark for Diffusion Models

GaussMarker combines spatial and frequency domain watermarking for enhanced robustness.

Key Features:

• Dual-domain watermarking

• Trained GNR (Gaussian Noise Residual) network

• High robustness to attacks

Usage:

watermark = AutoWatermark.load(
    'GM',
    algorithm_config='markdiffusion/config/GM.json',
    diffusion_config=diffusion_config
)

Training GNR Network:

python markdiffusion/watermark/gm/train_GNR.py --config markdiffusion/config/GM.json

PRC

Reference: An undetectable watermark for generative image models

PRC creates undetectable watermarks that are imperceptible to human observers and detection systems.

Key Features:

• Undetectable by design

• High security

• Minimal perceptual impact

Usage:

watermark = AutoWatermark.load(
    'PRC',
    algorithm_config='markdiffusion/config/PRC.json',
    diffusion_config=diffusion_config
)

SEAL

Reference: SEAL: Semantic Aware Image Watermarking

SEAL leverages semantic information to embed watermarks that adapt to image content.

Key Features:

• Semantic-aware embedding

• Content-adaptive watermarking

• Preserved semantic integrity

Usage:

watermark = AutoWatermark.load(
    'SEAL',
    algorithm_config='markdiffusion/config/SEAL.json',
    diffusion_config=diffusion_config
)

Video Watermarking Methods

VideoShield

Reference: VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking

VideoShield embeds watermarks into video generation models for content regulation and tracking.

Key Features:

• Video-specific watermarking

• Temporal consistency

• Frame-level detection

Usage:

watermark = AutoWatermark.load(
    'VideoShield',
    algorithm_config='config/VideoShield.json',
    diffusion_config=video_diffusion_config
)

# Generate watermarked video frames
frames = watermark.generate_watermarked_media(prompt)

# Detect in video
result = watermark.detect_watermark_in_media(frames)

VideoMark

Reference: VideoMark: A Distortion-Free Robust Watermarking Framework for Video Diffusion Models

VideoMark provides distortion-free watermarking specifically designed for video diffusion models.

Key Features:

• Distortion-free embedding

• Robust to video compression

• Temporal coherence preservation

Usage:

watermark = AutoWatermark.load(
    'VideoMark',
    algorithm_config='config/VideoMark.json',
    diffusion_config=video_diffusion_config
)

Algorithm Comparison

Choosing the Right Algorithm

For High Invisibility:

• Gaussian-Shading (GS) - Provably lossless

• PRC - Designed for undetectability

• Tree-Ring (TR) - Invisible in spatial domain

For High Robustness:

• ROBIN - Adversarial optimization

• GaussMarker (GM) - Dual-domain approach

• WIND - Two-stage robustness

For Video Content:

• VideoShield - Video regulation

• VideoMark - Distortion-free video watermarking

For Multi-User Scenarios:

• Ring-ID (RI) - Multi-key support

• SEAL - Semantic-aware adaptation

Algorithm Comparison

The following figure shows the performance comparison of different watermarking algorithms:

Next Steps

• ../tutorial - Hands-on tutorials for each algorithm

• Watermarking Workflow - Detailed watermarking workflow

• Visualization - Visualize how algorithms work

• Evaluation - Evaluate algorithm performance