Deep generative models excel at density estimation and distribution matching. However, their high fidelity is a double-edged sword. When trained on uncurated, real-world datasets, these models do not just learn the underlying semantic concepts; they also faithfully inherit and reproduce non-semantic artifacts, noise, and structural flaws embedded within the training data.
This phenomenon occurs because standard generative objectives treat all persistent statistical patterns equally. The model cannot inherently distinguish between semantic intent (e.g., the subject of a video) and statistical coincidence (e.g., a watermark overlay or compression artifact). As a result, the generated outputs inadvertently replicate the flaws of their training sets.
The Mechanism of Leakage
To understand why generative models memorize noise, consider how they learn a data distribution. In diffusion models, the training objective forces the network to learn a score function—the gradient of the log-data density. If a dataset contains systematic, persistent corruptions, the model perceives these corruptions as high-density regions of the true distribution.
As highlighted by Yangming Li et al. in Risk-Sensitive Diffusion: Robustly Optimizing Diffusion Models with Noisy Samples, failing to account for data corruption causes the model to learn inaccurate score functions, treating noise as an inherent signal. Similarly, Asad Aali et al. demonstrate in Solving Inverse Problems with Score-Based Generative Priors learned from Noisy Data that standard training objectives collapse when data is corrupted, forcing the model to represent the noise rather than the clean underlying signal. When the model later samples from this learned distribution, it reconstructs the artifacts along with the data.
Manifestations Across Modalities
Artifact inheritance manifests across a wide spectrum of generative tasks, ranging from high-level visual overlays to low-level statistical fingerprints.
Watermarks and Text Overlays
In large-scale web scraping, stock footage and branded images are ubiquitous. When video diffusion models are trained on these datasets, they frequently generate ghostly text or watermarks. For instance, Xiang Wang et al. observe this in VideoComposer: Compositional Video Synthesis with Motion Controllability, where persistent visual patterns from stock footage are baked into the synthesized output.
The same vulnerability appears across specialized video domains. Shenghai Yuan et al. note in MagicTime: Time-lapse Video Generation Models as Metamorphic Simulators that models fail to separate physical transformations from static text overlays. Even when structural consistency is improved, as in the cross-frame guidance framework by Jiasong Feng et al. (FancyVideo: Towards Dynamic and Consistent Video Generation via Cross-frame Textual Guidance), or the panoramic synthesis by Zixun Fang et al. (ViewPoint: Panoramic Video Generation with Pretrained Diffusion Models), watermarks and privacy-masking mosaics persist because the model treats them as inherent features of the data distribution.
Additional instances of this leakage are explored across various architectures and training frameworks, including:
- Runhui Huang et al. in DiffDis: Empowering Generative Diffusion Model with Cross-Modal Discrimination Capability
- Xinyuan Chen et al. in SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction
- Jinxi Xiang et al. in VersVideo: Leveraging Enhanced Temporal Diffusion Models for Versatile Video Generation
- Weiming Ren et al. in ConsistI2V: Enhancing Visual Consistency for Image-to-Video Generation
- Dejia Xu et al. in RAGME: Retrieval Augmented Video Generation for Enhanced Motion Realism
Low-Level Statistical Fingerprints
Artifact inheritance penetrates down to second-order statistics. Riccardo Corvi et al. demonstrate in Intriguing properties of synthetic images: from generative adversarial networks to diffusion models that generative models reproduce low-level JPEG compression signatures from their training sets. This creates a forensic paradox: synthetic images can masquerade as real compressed photos, misleading detection tools that rely on identifying compression signatures.
This low-level leakage spans modalities. In image enhancement, Yu-Sheng Chen et al. (Deep Photo Enhancer: Unpaired Learning for Image Enhancement From Photographs With GANs) show that models inherit haloing effects from traditional tone-mapping algorithms present in target datasets. In audio synthesis, Hang Zhou et al. (Vision-Infused Deep Audio Inpainting) note that models treating spectrograms as images replicate background noise from uncurated recordings rather than producing clean signals.
Suboptimal Design Mimicry
Beyond noise, models also inherit cognitive or procedural flaws. When generating structured data like architectural floor plans, training purely on existing distributions can be counterproductive. As Jun Yin et al. argue in FloorPlan-LLaMa: Aligning Architects' Feedback and Domain Knowledge in Architectural Floor Plan Generation, standard generative training causes models to replicate irrational design choices found in the source data, because the model lacks an inherent mechanism to distinguish high-quality architectural logic from poor execution.
Breaking the Inheritance Loop
To prevent models from treating artifacts as true signals, researchers are shifting away from naive distribution matching. One approach involves explicitly modeling the corruption process during training, allowing the network to mathematically isolate noise from the target distribution (e.g., using risk vectors or unbiased risk estimators). Another path introduces external feedback loops, such as reinforcement learning from domain knowledge or retrieval-augmented generation, to anchor the model's outputs in structural reality rather than uncurated dataset statistics.
Key Foundational Work
Deep Photo Enhancer: Unpaired Learning for Image Enhancement From Photographs With GANs (2018) — contributed 1 concept elements to this direction.
Vision-Infused Deep Audio Inpainting (2019) — contributed 1 concept elements to this direction.