🎬 AEGIS

AEGIS (Authenticity Evaluation on Generated vIdeo Samples) is a challenging benchmark that rigorously evaluates multimodal models’ capability in AI-generated video detection and reasoning.

This page hosts the AEGIS Hard Test Set, a carefully curated subset designed to stress-test model robustness under realistic, deceptive scenarios. For complete access to the full dataset including raw videos, features, and training splits, please refer to AEGIS-Full.

---

🔗 Related Datasets

• 🌐 AEGIS-Full The full dataset containing all splits (train, validation, hard test), raw clips, intermediate features, and annotations for exploring further interpretable synthetic video detection research.
• 📦 AEGIS-Baseline The training and validation subset used to reproduce baseline methods, including key frames and prompt-aligned annotations.

---

🧠 Pipeline Overview

The AEGIS Dataset Construction Pipeline: The process begins with data collection, sourcing both real videos and synthetic videos from advanced datasets and especially including self-generated highly fidelity videos. In data filtering, we rigorously remove low-quality or easily detectable samples. The curated set is then split into training, validation and a challenging hard test set to support robust evaluation. Finally, authenticity detection benchmarking are conducted with two prompting strategies.

---

📁 Data Format

Each example in the AEGIS Hard Test Set contains a rich set of multimodal features:

{
  "keyframes": ["Image", "Image", ..., "Image"],
  "optical_flow": ["Image", "Image"],
  "frequency_spectrum": "Image",
  "description": "Text summary of the video content",
  "reason": "Instruction guiding authenticity reasoning",
  "meta_data": "JSON-style string with label, resolution, etc.",
  "framediff_analysis": "Serialized JSON of per-frame diff metrics"
}

• 🖼 keyframes: 8 representative frames capturing the semantic core of the clip.
• 🌀 optical_flow: Two images representing temporal motion dynamics.
• 🌊 frequency_spectrum: Frequency domain signal of the first frame for structural artifact detection.
• 📝 description: Language-based summary of the video’s appearance.
• 💡 reason: Reasoning-oriented prompt tailored to test visual-linguistic comprehension.
• 📎 meta_data: Source, type (real/synthetic), resolution, and generation prompt metadata.
• 📊 framediff_analysis: Quantitative stats for motion inconsistency detection.

⚠️ Note: This subset contains a limited number of extracted features for demonstration and analysis. For full raw data and optical flow matrices, please refer to AEGIS-Full. A mapping file mapping_test_data.csv is also provided the link from video file names to unified indices.

---

📊 Baseline Evaluation Results

we provide our structured reasoning-prompt:

You are given a short video. Your task is to determine whether this video was generated by AI or captured from the real world.
You should think carefully and reason step-by-step based on the following aspects:

1. Frequency Domain Artifacts:
   Analyze each frame for signs of low high-frequency information, abnormal smoothness, or fingerprints left by generative models like diffusion models or GANs.
   AI-generated images often lack the fine details and sharpness found in real-world imagery.

2. Lighting and Shadow Consistency:
   Examine the interaction between light sources, shadows, and objects.
   In real scenes, lighting and shadow directions, sharpness, and intensity should be physically consistent.
   AI-generated frames often have subtle inconsistencies or unrealistic shading.

3. Compression Noise and Artifacts:
   Look for unnatural compression noise, grid-like block artifacts, or other signs of abnormal image processing.
   AI-generated videos sometimes introduce unique noise patterns different from natural camera compression.

4. Color, Texture, and Saturability:
   Inspect the colors and textures for unnatural smoothness, oversaturation, or inconsistencies in material appearance.
   Real-world scenes often have complex, imperfect textures that AI might oversimplify.

5. Physical Plausibility of the Scene:
   Evaluate whether the depicted objects, their arrangement, and their interactions are physically plausible.
   AI models sometimes generate scenes that subtly defy the laws of physics (e.g., impossible reflections, floating objects, inconsistent object deformations).

After careful analysis along these axes, decide whether the video is AI-generated or real.
Provide your reasoning based on the above factors.

YOU MUST FORMAT YOUR RESPONSE AS:
"AIGC Detection: [True/False]
Reasoning: [Your detailed reasoning]"

Zero-shot, Structured Prompt and LoRA Fine-tuning Evaluation on Hard Test Set

Model	Setting	Acc_all	Acc_real	Acc_ai	Macro F1
Qwen2.5-VL 3B	Zero-shot	0.52	0.80	0.23	0.48
Qwen2.5-VL 7B	Zero-shot	0.59	0.89	0.22	0.52
Video-LLaVA-HF 7B	Zero-shot	0.50	0.00	1.00	0.33
Qwen2.5-VL 3B	Reasoning-Prompt	0.47	0.58	0.35	0.46
Qwen2.5-VL 7B	Reasoning-Prompt	0.57	0.97	0.16	0.48
Video-LLaVA-HF 7B	Reasoning-Prompt	0.46	0.29	0.63	0.45
Qwen2.5-VL 3B	LoRA Fine-tuning	0.56	0.14	0.97	0.47
Qwen2.5-VL 7B	LoRA Fine-tuning	0.61	0.99	0.24	0.55

---

🙏 Acknowledgements

This dataset incorporates and builds upon the following sources:

• 📺 TIP-I2V dataset
• 🎥 Vript / Vript-Bench dataset
• 🧰 MM-Det synthetic video detection codebase

---

📜 License

Custom Academic Use License for AEGIS

By downloading, accessing, or using this dataset, you agree to the following terms:

1. Academic Use Only
   Strictly for non-commercial academic research purposes.

2. No Redistribution
   Respect the privacy of personal information of the original source. Without the permission of the copyright owner, you are not allowed to perform any form of broadcasting, modification or any other similar behavior to the data set content.

3. Attribution Required
   You must cite both this dataset (AEGIS) and the original source datasets it builds upon:

   • TIP-I2V: CC BY-NC 4.0 License
   • Vript: Academic use only, with prior permission.

4. No Warranty
   This dataset is provided "as-is", without any warranty or guarantees.

5. Revocation Clause
   Access and usage rights may be revoked in case of policy violations.

This license is adapted from HD-VG-100M and CC BY-NC 4.0 license frameworks.