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

Abstract: The rapid advancement of video generation models has enabled the creation of highly realistic synthetic media, raising significant societal concerns regarding the spread of misinformation. However, current detection methods suffer from critical limitations. They rely on preprocessing operations like fixed-resolution resizing and cropping. These operations not only discard subtle, high-frequency forgery traces but also cause spatial distortion and significant information loss. Furthermore, existing methods are often trained and evaluated on outdated datasets that fail to capture the sophistication of modern generative models. To address these challenges, we introduce a comprehensive dataset and a novel detection framework. First, we curate a large-scale dataset of over 140K videos from 15 state-of-the-art open-source and commercial generators, along with Magic Videos benchmark designed specifically for evaluating ultra-realistic synthetic content. In addition, we propose a novel detection framework built on the Qwen2.5-VL Vision Transformer, which operates natively at variable spatial resolutions and temporal durations. This native-scale approach effectively preserves the high-frequency artifacts and spatiotemporal inconsistencies typically lost during conventional preprocessing. Extensive experiments demonstrate that our method achieves superior performance across multiple benchmarks, underscoring the critical importance of native-scale processing and establishing a robust new baseline for AI-generated video detection.

• The paper demonstrates a native-scale Vision Transformer framework that preserves high-frequency forensic artifacts to robustly detect AI-generated videos.
• It employs dynamic resolution handling, 3D patchification, and fine-tuning strategies to achieve state-of-the-art accuracy on multiple benchmarks.
• Extensive evaluations show strong cross-model and cross-dataset transferability, validating the importance of updated, high-fidelity synthetic datasets.

Preserving Forgery Artifacts in AI-Generated Video Detection: A Native-Scale Vision Transformer Framework

Context and Motivation

The rapid proliferation of high-fidelity AI-generated videos, facilitated by the latest generation of diffusion, transformer-based, and commercial video synthesis models, imposes acute challenges for robust forgery detection. Previous detection methods are largely undermined by two factors: an overreliance on destructive preprocessing operations (such as fixed-resolution resizing and cropping, commonly to $224\times224$), and the use of outdated, low-quality synthetic datasets that no longer reflect the sophistication of contemporary generative engines. This produces detectors with poor generalization, primarily effective on obsolete artifacts or specific subsets of forgery patterns. The necessity for a scalable, robust, and semantically unbiased approach is amplified by the societal risks posed by synthetic media and the escalating realism of generated content.

Dataset Innovations

A central contribution of this study is a curated large-scale dataset containing over 140,000 videos spanning 15 open-source and proprietary generation models, complemented by the Magic Videos benchmark, which specifically targets scenarios susceptible to misuse (e.g., photorealistic landscapes, architecture, human interactions). The dataset ensures coverage of architectures ranging from Diffusion U-Net, DiT, and autoregressive models to state-of-the-art commercial engines (e.g., Sora, Jimeng, Pika Labs) with matched semantic content across real and fake distributions, thereby enabling unbiased, cross-model learning and evaluation.

Cross-validation experiments reveal strong positive correlation ($r = 0.86$) between the generation quality (as measured by established T2V benchmarks such as VBench) and the effectiveness of detectors trained on those data. Higher-quality synthetic data yields more transferable and robust forgery discriminative features, reaffirming the criticality of dataset curation and continual benchmark updates.

Methodological Contributions: Qwen2.5-ViT Native-Scale Detector

The detection framework leverages the Qwen2.5-VL Vision Transformer (ViT) as its visual backbone. Core architectural features include:

• Native-Scale 3D Patchification: Video inputs are partitioned into non-overlapping 3D patches at their native spatial and temporal scales (e.g., $2\times14\times14$). This eliminates the destructive effects of conventional resizing/cropping, fully preserving both high-frequency local artifacts (e.g., subtle texture errors, aliasing, unnatural edge transitions) and global semantic cues (e.g., object deformations, lighting inconsistencies).
• Dynamic Resolution Support: The model natively processes arbitrary video resolutions and durations, enabled by batch packing and hybrid windowed attention mechanisms, with practical GPU memory and training time optimizations through bfloat16 and FlashAttention.
• Fine-Tuning Strategies: Three approaches are explored—full fine-tuning, linear probing, and parameter-efficient LoRA—with full end-to-end backpropagation yielding the highest overall accuracy.

Experimental Analysis and Benchmark Results

Performance Overview

Extensive evaluations across multiple benchmarks (Magic Videos, MovieGen, DVF, GenVideo, DeepTraceReward) demonstrate that Qwen2.5-ViT achieves state-of-the-art accuracy and generalization, notably:

• Magic Videos (mACC): 83.2%—highest among tested detectors, including vision foundation models (CLIP, X-CLIP, Moon-ViT), image forgery detectors (NPR, FatFormer, B-Free), and deepfake detectors (F3Net, TALL).
• GenVideo-Val (F1): 90.64, AP: 96.13, bACC: 95.38—substantially outperforming DeMamba, RINE, and others, with only one-tenth the training data.
• DVF-Test (AUC): 97.6, and DeepTraceReward (ACC): 97.2%—highlighting robust cross-generator and cross-dataset transfer, especially on content from previously unseen engines.

Notably, the method maintains high balance between fake and real detection accuracies (96.3% vs. 98.2% on DeepTraceReward), avoiding the bias or overfitting common to detectors exposed to unbalanced or narrowly curated datasets.

Ablation Studies

• Spatial Resolution: Dynamic, aspect-ratio-preserving inputs up to 720p dramatically outperform fixed 224p cropping (mACC: 89.9 vs. 78.1), especially on high-fidelity synthetic data.
• Temporal Context: Increasing the temporal window (from 2 to 8 frames at 2 fps) enhances the model’s ability to capture spatiotemporal inconsistencies (88.6% vs. 82.9%).
• Robustness: The model is significantly less sensitive to minor compression (JPEG/H.264), downscaling, and cropping, maintaining strong detection accuracy under moderate perturbations and outperforming all baselines under severe spatial degradation.
• Efficiency: Despite a larger parameter count, Qwen2.5-ViT demonstrates favorable training efficiency (4.1x faster than CLIP-L, 25% lower GPU memory consumption), especially when using parameter-efficient fine-tuning.

Theoretical and Practical Implications

The principal theoretical implication is that native-scale, architecture-agnostic detectors are critical for robust generalization in AI-generated video detection. The findings empirically demonstrate that fixed-resolution downsampling, a ubiquitous preprocessing step in existing pipelines, causes irreversible loss of forensic cues and undermines model transferability to future, higher-quality synthesis models. Similarly, supervised learning on low-quality or outdated datasets produces detectors with little real-world robustness.

The practical implication is the necessity for continual dataset refreshes using the most sophisticated, semantically controlled synthetic content, and a shift toward dynamic-resolution, high-fidelity architecture paradigms. The Qwen2.5-ViT framework, with its architectural optimizations and native-scale 3D modeling, provides a scalable blueprint for future detection systems, balancing forensic fidelity with computational tractability.

Limitations and Future Directions

While the method delivers substantial gains, scalability to real-time or resource-constrained deployment remains a challenge due to the computational demands of native-resolution video processing. The dataset, while comprehensive, requires ongoing updates as generative modeling advances. Further, the explainability of the model’s decisions and identification of exact spatial-temporal regions responsible for classification remain unsolved—critical for forensics and legal contexts.

Future research directions include:

• Architectural efficiency enhancements for native-scale processing, potentially via sparse attention, feature selection, or early exit strategies.
• Automated, continuous dataset augmentation pipelines that can adapt to novel generative architectures and adversaries.
• Model explainability via interpretable attention mechanisms or localized saliency for forensic traceability.
• Integration with multimodal signals (e.g., audio, metadata) to further enhance detection in real-world media.

Conclusion

This work establishes that dynamic, native-resolution video modeling is essential for robust, generalizable detection of AI-generated video content, fundamentally surpassing the limitations imposed by legacy preprocessing and outdated datasets. By preserving crucial forgery artifacts and leveraging a highly optimized ViT-based architecture, the proposed framework sets a new baseline for AIGC detection, supporting both theoretical advancement and immediate practical application in digital forensics and information integrity.