Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models

2022/12/07 by Gowthami Somepalli, Somepalli, Gowthami, Vasu Singla +7 · 7 voices · 58 citations
Computer Science · Neuroscience · #Generative Adversarial Networks and Image Synthesis #Digital Media Forensic Detection #Aesthetic Perception and Analysis

paper · pdf · doi:10.48550/arxiv.2212.03860

Abstract

Cutting-edge diffusion models produce images with high quality and customizability, enabling them to be used for commercial art and graphic design purposes. But do diffusion models create unique works of art, or are they replicating content directly from their training sets? In this work, we study image retrieval frameworks that enable us to compare generated images with training samples and detect when content has been replicated. Applying our frameworks to diffusion models trained on multiple datasets including Oxford flowers, Celeb-A, ImageNet, and LAION, we discuss how factors such as training set size impact rates of content replication. We also identify cases where diffusion models, including the popular Stable Diffusion model, blatantly copy from their training data.

Cited by

• Dominant vs. Dominated: Concept-Level Generative Collapse in Diffusion Models
• Thinking in Video: Can Video Generators Really Reason About the Real World?
• Signed Rectified Flow: Negativity-Controlled Generation
• Addressing Limited Data in Auditory Attention Decoding with Diffusion Generative Models
• Points as Tori: Fast Pointwise Signed Distance for Point Clouds
• NAMESAKES: Probing Identity Memorization in Text-to-Image Models
• On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity
• Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge
• Memorization in 3D Shape Generation: An Empirical Study
• Taxing Artificial Intelligence
• GriDiT: Factorized Grid-Based Diffusion for Efficient Long Image Sequence Generation
• Metanetworks as Regulatory Operators: Learning to Edit for Requirement Compliance
• Copyright Infringement Risk Reduction via Chain-of-Thought and Task Instruction Prompting
• Beyond Memorization: Selective Learning for Copyright-Safe Diffusion Model Training
• CAPTAIN: Semantic Feature Injection for Memorization Mitigation in Text-to-Image Diffusion Models
• SUGAR: A Sweeter Spot for Generative Unlearning of Many Identities
• When Privacy Isn't Synthetic: Hidden Data Leakage in Generative AI Models
• Latent Diffusion Inversion Requires Understanding the Latent Space
• Generating Synthetic Human Blastocyst Images for In-Vitro Fertilization Blastocyst Grading
• GrOCE:Graph-Guided Online Concept Erasure for Text-to-Image Diffusion Models
• The Persistence of Cultural Memory: Investigating Multimodal Iconicity in Diffusion Models
• Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models
• On the Anisotropy of Score-Based Generative Models
• Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models
• Modal Aphasia: Can Unified Multimodal Models Describe Images From Memory?
• Gradient Variance Reveals Failure Modes in Flow-Based Generative Models
• DiffEM: Learning from Corrupted Data with Diffusion Models via Expectation Maximization
• Generative AI, UK Copyright and Open Licences: considerations for UK HEI copyright advice services
• Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive
• Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models
• Carré du champ flow matching: better quality-generalisation tradeoff in generative models
• How Diffusion Models Memorize
• Score-based Membership Inference on Diffusion Models
• Ethics Statements in AI Music Papers: The Effective and the Ineffective
• Copyright Infringement Detection in Text-to-Image Diffusion Models via Differential Privacy
• Review of Hallucination Understanding in Large Language and Vision Models
• No Prior, No Leakage: Revisiting Reconstruction Attacks in Trained Neural Networks
• A Unified Framework for Diffusion Model Unlearning with f-Divergence
• Responsible Diffusion: A Comprehensive Survey on Safety, Ethics, and Trust in Diffusion Models
• Latent Iterative Refinement Flow: A Geometric-Constrained Approach for Few-Shot Generation
• Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
• On the Edge of Memorization in Diffusion Models
• A Novel Metric for Detecting Memorization in Generative Models for Brain MRI Synthesis
• A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
• Mitigating data replication in text-to-audio generative diffusion models through anti-memorization guidance
• ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory
• AI-Generated Content in Cross-Domain Applications: Research Trends, Challenges and Propositions
• SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models
• On the Collapse Errors Induced by the Deterministic Sampler for Diffusion Models
• AMCR: A Framework for Assessing and Mitigating Copyright Risks in Generative Models
• Side Effects of Erasing Concepts from Diffusion Models
• Demystifying Foreground-Background Memorization in Diffusion Models
• SafeCtrl: Region-Based Safety Control for Text-to-Image Diffusion via Detect-Then-Suppress
• Score Augmentation for Diffusion Models
• Zero-Residual Concept Erasure via Progressive Alignment in Text-to-Image Model
• StyleSentinel: Reliable Artistic Copyright Verification via Stylistic Fingerprints
• LeakyCLIP: Extracting Training Data from CLIP
• LoReUn: Data Itself Implicitly Provides Cues to Improve Machine Unlearning

Discussions

• Literally, as in the math is literally just the reverse, it's just feeding image classification backwards This is why if you give it the description that was attached to an image in the training set, [bsky, 19 points, 1 comments]
• Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion [hn, 2 points, 0 comments]
• Have you read the literature in this space? This paper in particular shows that in many cases it’s not necessarily rare and that there often aren’t a lot of differences: https://arxiv.org/pdf/2212.038 [bsky, 2 points, 1 comments]
• Diffusion Art or Digital Forgery? Data Replication in Diffusion Models [hn, 2 points, 0 comments]
• The processes which humans learn by, and which GAN uses are not the same. A human is incapable of dissecting all this data and reassembling it via diffused dots generating hundreds finalized images in [bsky, 1 points, 0 comments]
• It’s a bit old now but there’s a study where you could regenerate the training data by prompting the keywords it’s tagged with Full article: arxiv.org/abs/2212.03860 [bsky, 1 points, 0 comments]
• Investigating Data Replication in Diffusion Models [hn, 1 points, 0 comments]

Related

• Image Copy Detection for Diffusion Models
• Measuring the Success of Diffusion Models at Imitating Human Artists
• Extracting Training Data from Diffusion Models
• Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images
• Replication in Visual Diffusion Models: A Survey and Outlook