Technical Summary: VINA - Video as Natural Augmentation for Unified AIGC Detection

1. Problem Statement

The rapid advancement of AI-Generated Content (AIGC) has created an urgent need for detectors that generalize across diverse data sources, deployment pipelines, and visual modalities. While state-of-the-art AI-generated image (AIGI) detectors perform competitively on image benchmarks, this paper identifies a critical and consistent failure mode: these detectors collapse when applied to frames extracted from AI-generated videos (AIGV).

The authors argue that treating AIGI and AIGV detection as separate tasks leaves a fundamental cross-modal generalization gap. This gap arises from two primary sources:

1. Synthesis-agnostic Video Processing Shifts: Real-world video frames undergo complex, entangled degradations not present in standard image datasets, including codec compression (e.g., H.264), motion blur, resizing, and limited-range color conversion (e.g., TV luminance ranges). These shifts suppress high-frequency micro-artifacts that image detectors often rely on as shortcuts.
2. Model-Specific Fingerprints: Modern video generators introduce distinct temporal artifacts and motion-related fingerprints that differ significantly from the spatial artifacts of image generators.

Consequently, detectors trained solely on images fail to recognize the specific statistical distributions and degradation patterns inherent in video frames, leading to performance drops approaching random guessing on video benchmarks.

2. Methodology: VINA

To address this gap, the authors propose VINA (VIdeo as Natural Augmentation), a unified framework that jointly trains on both image and video data to achieve robust cross-modal detection.

Core Philosophy

VINA treats video frames not merely as a separate task but as physically grounded natural augmentations for image data. Video frames expose detectors to realistic, complex degradations (compression, blur, color shifts) that are difficult to simulate with handcrafted 2D augmentations (e.g., Gaussian blur or JPEG re-compression).

Technical Components

1. Joint Training Paradigm: The framework trains a single detector on a mixed batch of static images ($D_{img}$) and video frames ($D_{vid}$). Video frames are modeled as pristine visual signals subjected to a complex, non-differentiable degradation function $T(\cdot)$ representing the video pipeline (codec, motion blur, resizing).
2. Cross-Modal Supervised Contrastive (CM-SupCon) Loss: To prevent the feature space from remaining modality-separated (where image and video features cluster separately despite sharing the same real/fake label), VINA introduces an explicit regularization term.
   • Mechanism: For a given anchor sample, the positive set includes only cross-modal samples (e.g., a video frame paired with an image) that share the same ground-truth label (real or fake).
   • Objective: This pulls real-image features toward real-video features and fake-image features toward fake-video features, while maintaining separation between real and fake classes. This forces the model to learn generative traces that are discriminative across both clean images and degraded video frames, suppressing reliance on modality-specific shortcuts.
3. Unified Objective: The final loss function combines standard binary cross-entropy ($L_{BCE}$) with the CM-SupCon term:
   $$L = L_{BCE} + \lambda L_{CM-SupCon}$$

3. Key Contributions

1. Identification of Systematic Failure: The paper reveals that SOTA AIGI detectors systematically fail on video benchmarks, formulating unified AIGC detection as a generalization problem that requires joint coverage of images and videos rather than isolated tasks.
2. VINA Framework: The proposal of a joint video-image training framework utilizing a CM-SupCon objective to regularize representations under a unified decision boundary, effectively suppressing source- and modality-specific shortcut cues.
3. Comprehensive Evaluation: Extensive experiments across 14 benchmarks (including image, video, and in-the-wild datasets) demonstrate that VINA achieves bidirectional gains, improving performance on both video and image tasks without complex dataset-specific tuning or explicit alignment pipelines.

4. Experimental Results

The authors evaluated VINA against a wide range of baselines, including video-specific detectors (e.g., DeMamba, WaveRep), image-specific detectors (e.g., NPR, Effort, Co-SPY), and methods trained on aligned datasets (e.g., DDA, B-Free).

• Benchmark Performance:
  • Video Benchmarks: VINA achieved an average accuracy of 90.4%, significantly outperforming image-only detectors (which often dropped below 60%) and competitive video-specific detectors.
  • Image Benchmarks: VINA achieved 94.1% average accuracy, surpassing specialized image detectors and demonstrating that video data does not degrade image performance.
  • In-the-Wild Benchmarks: On challenging real-world datasets (e.g., Chameleon, SynthWildX, WildRF), VINA reached 92.9% average accuracy, showing superior robustness to unknown generators and post-processing compared to prior art.
• Robustness: The model demonstrated strong resilience to various perturbations, including JPEG/HEIF compression, scaling, and cropping, without requiring explicit format alignment during training.
• Efficiency: Despite using a large DINOv3 backbone, VINA operates at 256×256 resolution with 82.4G FLOPs per frame, achieving the fastest inference throughput (15.7s per 1K frames) among compared methods.
• Ablation Studies:
  • Backbones: Gains were consistent across CLIP, DINOv2, and DINOv3 backbones.
  • Data Sources: Performance improvements were observed across various data combinations (e.g., Pika+ProGAN, GenVideo+SD1.4, 140k+DDA), indicating the method is not tied to specific datasets.
  • Augmentation vs. Joint Training: Synthetic degradation strategies (color jitter, HEIF compression) provided moderate improvements but were significantly outperformed by using real video data as natural augmentation.

5. Significance and Claims

The paper claims that VINA establishes a new paradigm for AIGC detection by treating realistic video degradations as a form of natural augmentation rather than a separate domain. The significance lies in:

• Unified Detection: Demonstrating that a single model can effectively detect both images and videos, bridging the cross-modal gap that currently plagues the field.
• Robustness without Complexity: Achieving state-of-the-art performance without relying on complex data alignment, heavy augmentation pipelines, or dataset-specific hyperparameter tuning.
• Generalization: Providing a solution that generalizes across unseen generators and in-the-wild scenarios, addressing the critical need for detectors that function reliably in diverse real-world deployment settings.

The authors conclude that while their current detector operates on single frames and does not explicitly model temporal relationships, the joint training philosophy effectively leverages the physical properties of video data to create a more robust and universal detection framework.