Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning

Technical Summary: Skyra – AI-Generated Video Detection via Grounded Artifact Reasoning

Problem Statement
The rapid advancement of diffusion-based and multimodal generative models has enabled the creation of photorealistic synthetic videos, posing significant risks to social safety through misinformation and deepfakes. While existing detection methods have largely focused on binary classification using spatio-temporal features, they suffer from two critical limitations: a lack of interpretability (opaque decision processes) and an inability to capture human-perceivable artifacts. Empirical analysis in the paper reveals that even state-of-the-art general Multimodal Large Language Models (MLLMs) perform near-randomly on this task, often failing to identify subtle generative artifacts or misinterpreting natural video degradations (e.g., motion blur, compression) as forgery. Furthermore, existing MLLM-based detectors tend to rely on superficial cues like visual quality or lighting rather than intrinsic, physics-violating inconsistencies that humans use to identify synthetic content.

Methodology
The authors propose Skyra, a specialized MLLM designed to detect AI-generated videos by identifying and reasoning about grounded visual artifacts. The methodology is built upon three core components:

1. ViF-CoT-4K Dataset: To address the lack of fine-grained human annotations, the authors constructed the first large-scale dataset of AI-generated video artifacts. This dataset includes:

   • Data Collection: Real videos from sources like Panda-70M and Kinetics-400, and synthetic videos generated by over ten state-of-the-art models (e.g., Sora-2, Wan2.2, Kling, HunyuanVideo).
   • Hierarchical Taxonomy: A three-layer artifact taxonomy distinguishing between Low-level Forgery (e.g., texture anomalies, color/lighting issues) and Violation of Laws (e.g., object inconsistency, unnatural movement, causality violations).
   • Annotation Process: Professional annotators performed side-by-side comparisons of real and fake videos to identify artifacts, providing temporal-spatial localization (timestamps and bounding boxes) and textual explanations. These were converted into Chain-of-Thought (CoT) data using a prompt-engineering strategy involving self-curation and in-context learning.

2. Two-Stage Training Strategy:

   • Stage 1: Supervised Fine-Tuning (SFT) / Cold-Start Initialization: The base model (Qwen2.5-VL-7B) is fine-tuned on ViF-CoT-4K using a structured output format that requires the model to output a <thinking> block (reasoning process) and an <answer> block (verdict). For fake videos, the model must explicitly tag artifacts with type, time, and location; for real videos, it must systematically inspect and clear suspected regions. This stage endows the model with basic detection and explanation capabilities.
   • Stage 2: Reinforcement Learning (RL): To enhance the model's ability to discover discriminative artifacts and improve generalization, the authors employ Group Relative Policy Optimization (GRPO). A custom reward function is designed with two components:
     • Asymmetric Accuracy Reward: Penalizes false positives (predicting "Fake" for real videos) more severely than false negatives to prevent bias toward the "Fake" class.
     • Inspection Reward: Encourages the model to adhere to the required format and actively explore visual cues, rather than relying on superficial patterns.

3. ViF-Bench Benchmark: A comprehensive evaluation benchmark comprising 3,000 high-quality samples generated by over ten recent video generators, aligned in semantics and format with real counterparts to mitigate shortcut learning.

Key Contributions

• Skyra Model: A specialized MLLM that achieves state-of-the-art performance in both detection accuracy and explainability by leveraging grounded artifact reasoning.
• ViF-CoT-4K Dataset: The first large-scale dataset featuring fine-grained, human-annotated spatio-temporal artifacts with a hierarchical taxonomy, enabling supervised training for explainable detection.
• ViF-Bench: A rigorous benchmark designed to evaluate detection robustness across diverse, state-of-the-art generators, featuring real-fake pairs with aligned semantics.
• Training Framework: A novel two-stage pipeline combining SFT with a carefully designed RL strategy (including asymmetric rewards) to improve artifact perception and reasoning capabilities.

Experimental Results
Extensive experiments on ViF-Bench and the external GenVideo benchmark demonstrate Skyra's superiority:

• Detection Performance: Skyra-RL achieves an accuracy of 91.02% and an F1-score of 90.27% on ViF-Bench, outperforming the second-best method (DeMamba) by +17.27% in F1-score and surpassing existing MLLM-based detectors by a significant margin (e.g., +32% F1 over BusterX++).
• Generalization: Skyra shows strong cross-domain generalization. When adapted to the GenVideo benchmark (out-of-domain), it achieves a +19.22% accuracy gain over its SFT-only version, indicating that RL helps the model learn intrinsic artifact cues rather than generator-specific patterns.
• Robustness: The model maintains high performance under various video degradations, including compression, Gaussian noise, and color transformations, where binary classifiers and other MLLMs often fail.
• Explanation Quality: Case studies show that Skyra can identify subtle, human-perceivable artifacts (e.g., shape distortion, abnormal object appearance) and provide coherent, grounded explanations with precise spatio-temporal localization, whereas baseline MLLMs often provide generic descriptions or miss artifacts entirely.

Significance and Claims
The paper claims that Skyra represents a shift from opaque binary classification to interpretable, artifact-centric detection. By mimicking human reasoning—actively searching for physical and logical inconsistencies—the model provides not just a verdict but a "grounded" explanation that can assist human verification. The authors posit that their dataset and benchmark offer a standardized testbed for advancing the field of explainable AIGC detection. They acknowledge limitations, noting that the model's performance is bound by the generators covered in their dataset and that it does not assess the intent or societal harm of a video. They emphasize that Skyra is designed to support human-in-the-loop verification rather than replace it, highlighting the need for calibrated uncertainty estimation in safety-critical applications.