TL;DR
CreaTrace is a new framework designed to detect "creative plagiarism"—where the narrative structure and shot design of a video are copied even if the visual appearance (actors, settings, etc.) is entirely different. By decoupling abstract creative patterns from specific visual realizations and using an iterative evidence-tracing approach, it significantly outperforms traditional video copy detection methods and state-of-the-art multimodal large language models (MLLMs).
What Problem Does This Paper Address?
Traditional video copyright protection focuses on Video Copy Detection (VCD)—finding instances where the actual pixels or visual content have been reused or slightly modified (e.g., cropped, filtered, or re-encoded).
However, the rise of short-video platforms and generative AI has birthed a more subtle challenge: Video Creative Plagiarism. This occurs when a creator reproduces the "creative design" of a video—its narrative arc, specific conflict ideas, and shot arrangements—while changing the visual "realization." For example, one video might show a factory worker buying a steamed bun (baozi), while another shows a laborer buying beef noodles, but both follow the exact same sequence: price rise $\rightarrow$ protest $\rightarrow$ new store opens $\rightarrow$ customers flock to the new stall.
Existing methods fail here because they are too "pixel-hungry"; they look for visual similarity rather than structural or narrative similarity. The authors introduce the task of Video Creative Plagiarism Detection (VCPD) to bridge this gap.
Methods
The authors propose CreaTrace, a framework that moves beyond visual matching through several novel stages:
- Multimodal Semantic Event Segmentation (MSES): Instead of simple shot-based cutting, MSES uses visual cues, speech transcripts (via WhisperX), and MLLMs (Qwen3-VL) to divide long videos into coherent "narrative events." This provides a structured foundation for analysis.
- Coarse Creative Matching: An initial filter that compares event semantics (using BGE embeddings) and "camera language" (shot types and editing rhythms) using Dynamic Time Warping.
- Creative Decoupling: This is the core innovation. Using a dual-query Q-Former, the model learns to separate a video's features into two latent spaces: Creative (the abstract pattern) and Realization (the specific visual look).
- Creative Swap Contrastive Learning: To ensure the "Creative" features are truly independent of the visuals, the model performs a "swap." It takes the creative features of Video A and the realization features of Video B to reconstruct a hybrid. It then uses contrastive loss to ensure that the creative identity remains consistent regardless of which "visual body" it inhabits.
- Iterative Evidence Tracing: The model doesn't just give a "yes/no" answer. It iteratively refines the localization of suspicious segments, feeding localized evidence back into the matching process to improve accuracy and generate an interpretable "Evidence Report."
Datasets
Because VCPD is a new task, the authors constructed a specialized dataset:
- Source: 40 publicly available videos spanning 10 "creative groups" (groups of videos that share the same underlying idea but different visuals).
- Scale: 17.82 hours of footage, segmented into 515 semantic events.
- Pairs: They generated over 22,000 positive event pairs (within the same creative group) and 103,000 negative pairs.
- Evaluation Set: A balanced test set of 80 video pairs (40 positive, 40 negative) from creative groups not seen during training.
Key Insights and Ideas
- Creativity is Structural, Not Just Semantic: The paper argues that knowing what is in a video (semantics) isn't enough; you must understand how it is organized and presented (structure and camera language).
- The "Swap" Trick: By forcing the model to maintain creative consistency across swapped visual realizations, the authors effectively "blind" the model to superficial visual differences, forcing it to focus on the underlying narrative "skeleton."
- Interpretability Matters: Unlike black-box detectors, CreaTrace produces an evidence report that maps specific narrative correspondences (e.g., "Stock Shortage $\rightarrow$ Long Wait"), which is crucial for actual legal or platform-level copyright enforcement.
Key Results
CreaTrace was compared against zero-shot MLLMs (like Qwen3-VL and InternVL3) and traditional copy localization methods (like TransVCL):
- Detection Accuracy: CreaTrace achieved an Accuracy (ACC) of 68.75% and an Average Precision (AP) of 90.48%, significantly outperforming the best-performing MLLM (Qwen3-VL), which hit 52.50% ACC and 62.95% AP.
- Localization: In evidence localization (finding the exact timestamps of the plagiarism), CreaTrace achieved an F1-score of 30.19 at a tIoU threshold of 0.5, while traditional methods like TransVCL and RTR failed almost entirely (scoring near 0 or 4.55) because they couldn't handle the visual domain shift.
- Iterative Refinement: The authors showed that using the localized evidence to "re-query" the model (Iteration 2) improved the performance of even baseline MLLMs by over 20-30%.
Limitations and Future Work
- Dataset Scale: While the dataset is high-quality, 40 source videos is a small sample size for the vast diversity of internet content. Expanding this to more genres (e.g., dance, tutorials) is a likely next step.
- Computational Overhead: The framework relies on multiple passes of MLLMs and encoders. For real-time monitoring on a platform like TikTok or YouTube, the inference speed would need significant optimization.
- Defining "Creativity": The paper focuses on narrative and shot design. Future work might explore other creative dimensions, such as color grading styles or specific comedic timing patterns.
Summary updated