Preserving Forgery Artifacts: AI-Generated Video Detection at Native Scale

1. Problem Statement

The rapid advancement of AI video generation models (e.g., Sora, Kling, Wan2.1) has led to the creation of hyper-realistic synthetic media, posing significant risks of misinformation and deepfakes. Current detection methods suffer from two critical limitations:

• Destructive Preprocessing: Existing detectors typically resize or crop input videos to a fixed low resolution (e.g., 224×224). This process discards subtle, high-frequency forgery artifacts (pixel-level traces) and causes spatial distortion, leading to a loss of critical forensic cues.
• Outdated Data: Most benchmarks rely on datasets generated by older, lower-quality models. Detectors trained on these fail to generalize to modern, high-fidelity generators.
• Resolution Sensitivity: Preliminary experiments show that detectors trained on specific resolutions suffer significant performance drops when evaluated on videos of different resolutions, and performance is positively correlated with the quality of the training generator.

2. Methodology

The authors propose a unified framework that operates natively on variable spatial resolutions and temporal durations, avoiding the information loss associated with downsampling.

A. Data Curation

• Training Dataset: A large-scale dataset comprising 140,000+ videos generated by 15 state-of-the-art models (including Diffusion U-Nets, Diffusion Transformers, and Auto-regressive models) and an equal number of real videos from Kinetics-400 and MSVD.
• Magic Videos Benchmark: A new evaluation benchmark designed specifically for ultra-realistic content. It uses a curated prompt library to generate high-quality synthetic videos from 6 recent generators (e.g., Wan2.1, Hailuo, StepVideo) and pairs them with real videos from Mixkit and Pexels.
• Prompt Strategy: Real video captions are refined using GPT-4o to create consistent, high-quality prompts, ensuring the synthetic videos are indistinguishable from real content in terms of semantics, isolating the forgery artifacts as the primary detection signal.

B. Model Architecture: Qwen2.5-ViT

The core of the detection framework is built upon the Qwen2.5-VL Vision Transformer.

• Native-Scale 3D Patchification: Instead of resizing frames, the model tokenizes input videos into 3D patches (spatial × temporal) at their original resolution. For a video tensor $V \in \mathbb{R}^{T \times H \times W \times C}$, it partitions the data into non-overlapping 3D patches of size $(2, 14, 14)$.
• Preservation of Artifacts: By maintaining the native aspect ratio and resolution, the model preserves high-frequency textures and spatiotemporal inconsistencies that are typically destroyed by resizing.
• Efficiency Optimizations: To handle the computational cost of high-resolution inputs, the framework employs:
  • Batch Packing (NaViT): Handles variable-length sequences without padding.
  • Flash Attention: Optimizes memory and compute for long sequences.
  • Windowed Attention: Uses a hybrid attention strategy (114×114 windows) to ensure linear scaling of computational cost.
• Training Strategy: The model supports Full Fine-tuning, Linear Probing, and Parameter-Efficient Fine-Tuning (LoRA).

3. Key Contributions

1. High-Quality Diverse Dataset: Introduction of a 140K+ video training set from 15 generators and the "Magic Videos" benchmark, ensuring alignment with the current frontier of AIGC quality.
2. Native-Resolution Framework: A novel detection architecture based on Qwen2.5-ViT that processes videos at their native aspect ratio and resolution, eliminating the need for destructive resizing/cropping.
3. State-of-the-Art Performance: The method establishes a new baseline for AI-generated video detection, demonstrating superior generalization across diverse generators and resolutions compared to existing SOTA methods.

4. Experimental Results

The proposed method (Qwen2.5-ViT) was evaluated on multiple benchmarks:

• Magic Videos Benchmark:

  • Achieved 97.20% Accuracy (ACC) and 93.28% Average Precision (AP) on average.
  • Significantly outperformed specialized deepfake detectors (e.g., F3Net, TALL) and general image detectors (e.g., NPR, Effort), which struggled with the temporal dynamics and high resolution of modern videos.
  • Showed robust performance across all 6 test generators, with the highest mACC among all baselines.

• Cross-Dataset Evaluation:

  • DVF Test: Achieved 97.6% AUC, outperforming all baselines (next best was 95.4%).
  • GenVideo-Val: Achieved 96.64% ACC and 96.13% AP, outperforming DeMamba (trained on 2.2M samples) despite being trained on only 140K samples.
  • DeepTraceReward: Achieved 97.2% ACC, significantly surpassing large Multimodal LLMs (e.g., GPT-5, Gemini 2.5 Pro) which struggled with the binary classification task.

• Ablation Studies:

  • Resolution: Dynamic resolution (up to 720p) yielded 89.92% average accuracy, compared to 73.69% for random resizing to 224p. This confirms that preserving native resolution is critical for detecting subtle artifacts.
  • Temporal Context: Increasing sampled frames from 2 to 8 improved performance from 82.93% to 88.60%.
  • Robustness: The model remained highly accurate under moderate JPEG and H.264 compression, outperforming baselines even under aggressive downscaling.

5. Significance

• Paradigm Shift: The paper challenges the industry standard of fixed-resolution preprocessing, demonstrating that "native-scale" processing is essential for detecting sophisticated AI forgery.
• Generalization: By training on diverse, high-quality data without resolution constraints, the model learns generalizable forgery features rather than overfitting to specific generator artifacts or resolution patterns.
• Practical Impact: The framework provides a robust defense against the next generation of deepfakes, which are increasingly photorealistic and generated at high resolutions. It sets a new standard for future research in AIGC detection, emphasizing the importance of data quality and architectural fidelity to the input signal.