PaperlayerView exact revision on arXiv

Source-linked AI summary

Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning

Yifei Li, Wenzhao Zheng, Yanran Zhang, Runze Sun, Yu Zheng, Lei Chen, Jie Zhou, Jiwen Lu

arXiv:2512.15693v2cs.CV

TL;DR

AI-generated video detectors often provide only binary decisions, while general MLLMs perform near randomly and miss human-perceivable artifacts. Skyra addresses this gap with artifact-grounded reasoning, fine-grained annotations, and two-stage training, outperforming existing detectors across benchmarks while producing interpretable explanations.

  • Problem

    Existing AI-generated video detectors are largely limited to binary classification, lack detailed rationales, and general MLLMs perform near randomly while missing human-perceivable artifacts.

  • Method

    Skyra uses a specialized MLLM with a hierarchical artifact taxonomy, the human-annotated ViF-CoT-4K dataset, and supervised initialization followed by reinforcement learning.

  • Results

    Skyra substantially outperforms existing binary and MLLM-based detectors across ViF-Bench and GenVideo, including +26.73% accuracy and +17.27% F1 over the second-best DeMamba.

  • Takeaways & Limitations

    Grounded artifact reasoning enables Skyra to combine AI-generated video detection with coherent, human-readable explanations and spatio-temporal artifact perception.

  • Takeaways & Limitations

    The benchmark does not cover all emerging media distributions, and Skyra does not assess intent, context, or societal harm; its rationales may be overconfident or partially hallucinated.

Abstract

from arXiv · show

The misuse of AI-driven video generation technologies has raised serious social concerns, highlighting the urgent need for reliable AI-generated video detectors. However, most existing methods are limited to binary classification and lack the necessary explanations for human interpretation. In this paper, we present Skyra, a specialized multimodal large language model (MLLM) that identifies human-perceivable visual artifacts in AI-generated videos and leverages them as grounded evidence for both detection and explanation. To support this objective, we construct ViF-CoT-4K for Supervised Fine-Tuning (SFT), which represents the first large-scale AI-generated video artifact dataset with fine-grained human annotations. We then develop a two-stage training strategy that systematically enhances our model's spatio-temporal artifact perception, explanation capability, and detection accuracy. To comprehensively evaluate Skyra, we introduce ViF-Bench, a benchmark comprising 3K high-quality samples generated by over ten state-of-the-art video generators. Extensive experiments demonstrate that Skyra surpasses existing methods across multiple benchmarks, while our evaluation yields valuable insights for advancing explainable AI-generated video detection.

1. Introduction

AI-generated videos are increasingly realistic and socially risky, while existing detectors often provide binary decisions without human-interpretable artifact evidence. Skyra addresses this gap as a specialized MLLM for grounded artifact-based detection and explanation.

  • Synthetic videos have become highly realistic, increasing concerns about their misuse and social safety.
  • Existing MLLM detectors can perform near randomly, miss human-perceivable artifacts, and overemphasize superficial cues.
  • Skyra is introduced as a specialized MLLM for AI-generated video detection via grounded artifact reasoning.

Skyra

Skyra identifies spatio-temporally grounded artifacts and uses them to support both detection and human-interpretable explanations. It outperforms binary classifiers and prior MLLM-based approaches in detection performance.

  • Skyra demonstrates superior artifact perception and detection capabilities compared with off-the-shelf MLLMs and previous MLLM-based detectors.
  • Skyra identifies artifacts as spatio-temporally grounded evidence for AI-generated video detection and explanation.
  • Skyra provides fine-grained, human-interpretable artifact localization while outperforming binary classifiers and prior MLLM-based detectors.

2. Related Work

Prior AI-generated video detectors largely rely on binary classification and lack interpretability, while video MLLMs and reinforcement learning provide relevant foundations for more explainable reasoning. Existing work motivates artifact-focused supervision and post-training for video detection.

  • AI-Generated Video Detection: Binary AI-generated video detectors can perform well on benchmarks but leave their detection process opaque and difficult to verify manually.
  • Multimodal Large Language Models for Video: MLLM research has enabled reasoning-based detection in faces and images, but video-specific artifact supervision remains a distinct need.
  • Multimodal Large Language Models for Video: Video MLLMs support spatiotemporal reasoning, while reinforcement learning is used to address limited reasoning in base MLLMs.

3. ViF Dataset

ViF-CoT-4K and ViF-Bench are designed to address shortcut-prone datasets, limited generator diversity, and missing artifact rationales. The dataset combines a hierarchical artifact taxonomy, diverse generators, and fine-grained manual annotation.

  • Figure 4 summarizes generator distributions, artifact-type distributions, and the CoT annotation vocabulary in ViF-CoT-4K and ViF-Bench.
  • Existing datasets suffer from real-fake discrepancies, limited generator diversity, and a lack of detailed artifact annotations.
  • The dataset and benchmark construction pipeline uses fine-grained manual annotations to address these limitations.
  • Artifact Taxonomy: The hierarchical taxonomy organizes human-perceivable artifacts into three layers, from broad forgery categories to fine-grained observable cues.
  • Training and evaluation use diverse open-source and commercial video generators, including recent models such as Wan2.2, LTX-Video, MiniMax-Hailuo, and Sora-2.

4. Skyra

Skyra adapts MLLMs for AI-generated video detection through two-stage training that combines supervised reasoning formats with reinforcement learning for grounded artifact discovery.

  • 4. Skyra: Skyra uses supervised fine-tuning followed by reinforcement learning to develop basic detection and explanation abilities, then improve artifact perception and reasoning.The supervised stage trains structured responses, while RL uses redesigned task rewards to elicit self-coherent forgery-cue discovery.
  • 4. Skyra: Human inspection motivates searching for spatio-temporal inconsistencies, but off-the-shelf MLLMs often miss subtle forgery cues and mistake natural degradations for evidence.Direct prompting and chain-of-thought strategies often achieve below 60% accuracy on the authors’ benchmarks.
  • 4. Skyra: The supervised format requires a Fake or Real answer with explicit reasoning, while fake predictions additionally identify forgery type and grounded temporal-spatial regions.Real-video responses use the same temporal-spatial tags to inspect suspected regions.
  • 4. Skyra: The SFT stage fine-tunes Qwen2.5-VL-7B on ViF-CoT-4K using autoregressive cross-entropy over annotated response sequences.The paper reports that this cold-start initialization is crucial because without it, forgery recognition remains weak and later RL receives sparse rewards.
  • 4. Skyra: RL combines classification accuracy and format-validity rewards, weighting them 0.8 and 0.2 while penalizing false positives asymmetrically.The reward design reflects that detecting Fake requires one artifact, whereas confirming Real requires broader evidence.

5. Experiments

Skyra consistently outperforms binary detectors and MLLM baselines across in-domain, out-of-domain, and degraded-video evaluations. Ablations and case studies indicate that chain-of-thought reasoning, supervised cold-start training, and reinforcement learning support detection and artifact-focused explanations.

  • 5.2. Main Results: +26.73% accuracy and +17.27% F1 over DeMamba establish Skyra as the strongest model on ViF-Bench.Against MLLM baselines, Skyra also gains +34.12% accuracy, +24.57% recall, and +32% F1.
  • 5.2. Main Results: +11.07% accuracy over the best binary detector demonstrates stronger out-of-domain performance on GenVideo.Skyra also improves over Skyra-SFT by +7.8% accuracy and +16.9% recall.
  • 5.2. Main Results: +19.22% accuracy, +42.06% recall, and +31% F1 result from one-epoch RL adaptation using 2.2K GenVideo samples without additional human annotation.The adaptation starts from Skyra-SFT and follows GenVideo’s many-to-many setting.
  • 5.2. Main Results: Skyra maintains state-of-the-art performance under compression, zoom, Gaussian noise, lighting, and color transformations.These five degradation types are intended to reflect perturbations in degraded real-world videos.
  • 5.3. Ablation Study: Removing chain-of-thought reasoning yields near-random performance, while removing cold-start initialization performs even worse; reinforcement learning further improves the supervised model.The ablation also reports suboptimal performance from direct binary rewards.
  • 5.4. Case Study: Skyra’s explanations can identify fine-grained artifacts, while real-video responses follow a description–inspection–review–conclusion process.The case study presents artifact perception as a result of cold-start initialization and reinforcement learning.

6. Conclusion and Discussions

The paper concludes that Skyra combines artifact-centric detection with grounded explanations and strong benchmark performance. It also identifies coverage, deployment, and rationale-calibration boundaries that limit current applicability.

  • Conclusion: Skyra provides interpretable artifact-centric detection through spatio-temporally grounded evidence and coherent explanations.The system is built with ViF-CoT-4K and a two-stage supervised-initialization and reinforcement-learning pipeline.
  • Limitations: The benchmark does not encompass all emerging media distributions, including ultra-long videos and non-photorealistic stylized content.The training data remain tied to the generators and collection pipeline used for ViF-CoT-4K and ViF-Bench.
  • Limitations: Skyra does not assess a video’s intent, context, or potential societal harm.Its human-readable rationales may still be overconfident or partially hallucinated, motivating calibrated uncertainty, human oversight, and complementary safeguards.
  • Artifact Taxonomy: The artifact taxonomy covers low-level forgery cues, object and interaction inconsistencies, unnatural movement, causal violations, texture and blur anomalies, color and lighting anomalies, and camera-motion inconsistencies.Violation-of-laws artifacts require spatio-temporal reasoning and background knowledge about real-world behavior.

B. Analysis of Generalization Ablations.

The ablation shows that reinforcement learning is the primary driver of Skyra’s cross-domain generalization, recovering performance lost to SFT overfitting while preserving in-domain accuracy.

  • B. Analysis of Generalization Ablations.: +7.8% cross-dataset accuracy over SFT is achieved by the subsequent RL stage while in-domain performance is preserved.SFT raises in-domain accuracy but reduces cross-dataset accuracy from 68.34% to 63.98%.

C. Analysis of Artifact Cues

Skyra’s detected artifacts are dominated by semantic violations of physical and logical consistency rather than low-level visual forgery cues.

  • C. Analysis of Artifact Cues: 82.8% of detected artifacts are Violation of Laws, compared with 17.2% Low-Level Forgery, indicating semantic rather than low-level artifact reasoning.The distribution is reported in Table 7.
  • C. Analysis of Artifact Cues: Interaction Inconsistency accounts for 10.0%, while remaining categories account for 44.7%, spanning object permanence, causality, and commonsense.These categories show that Skyra attends to diverse physical violations.
  • C. Analysis of Artifact Cues: The 17.2% Low-Level Forgery share suggests Skyra avoids overfitting to generator-specific texture and color cues.The analysis links semantic-level violations to stronger cross-generator and cross-dataset generalization.

D.1. Binary Detectors

The reviewed binary and MLLM-based detectors exhibit limited robustness or interpretability: they may overfit distributions, attend to irrelevant regions, or miss localized forgery artifacts.

  • D.1. Binary Detectors: Demamba’s fake and real embeddings overlap heavily, while CAMs focus on fixed spatial locations rather than authenticity-discriminative content.The analysis indicates sensitivity to textures outside the main frame content and limited generalization.
  • D.1. Binary Detectors: NSG-VD achieved near-perfect AU-ROC on validation but performed poorly on test data, indicating strong overfitting.Reusing the training-time reference dataset may partly explain its suboptimal OOD and robustness performance.
  • D.1. Binary Detectors: NSG-VD’s reference-dataset mechanism limits effectiveness when target real samples are unavailable or should not enter the reference set.This constrains performance in OOD or noisy scenarios.
  • D.2. Off-the-Shelf MLLMs: Off-the-shelf MLLMs confidently misclassify synthetic videos as real when they describe overall semantics instead of subtle spatiotemporal artifacts.They can also mistake compression, blur, and low-light noise in real videos for evidence of forgery.
  • D.2. Off-the-Shelf MLLMs: BusterX++ succeeds when synthetic content has obvious global stylistic discrepancies but misses localized physics violations despite coherent overall appearance.Its behavior illustrates the limitation of relying primarily on global scene appearance.

F. Broader Impacts

Skyra is motivated by societal risks from AI-generated videos and aims to support human verification through artifact-grounded explanations and standardized evaluation resources, while acknowledging dual-use risks.

  • F. Broader Impacts: AI-generated videos threaten social trust through misinformation, impersonation, and erosion of confidence in authentic media.The paper frames these risks as motivation for interpretable detection.
  • F. Broader Impacts: Skyra’s artifact-centric outputs are intended to help journalists, fact-checkers, regulators, and platform moderators assess suspicious content.The system is designed to provide grounded visual evidence alongside predictions.
  • F. Broader Impacts: ViF-CoT-4K and ViF-Bench provide a standardized testbed for evaluating detectors on diverse, up-to-date generators.The paper presents these resources as supporting more reliable and transparent AIGC safety tools.
  • F. Broader Impacts: The artifact taxonomy, benchmark, and detector create dual-use concerns because detector failure-mode insights could inform more evasive generative models.The authors state that the defensive benefits outweigh these risks and describe mitigation through dataset curation.

G. License

The supplied material documents ViF-CoT-4K’s artifact annotation framework, example outputs, benchmark and generator references, and dataset licensing information.

  • G. License: ViF-CoT-4K and ViF-Bench are released under CC BY 4.0, with users also required to follow the licenses of the source datasets.
  • G. License: The annotation framework requires separate fake and real video CoT annotations in a single JSON object with prescribed output formats.
  • G. License: Fake-video reasoning must identify all annotated artifacts in temporal order, while real-video reasoning must inspect and clear corresponding regions.
  • G. License: Examples ground artifacts with category labels, time intervals, and spatial bounding boxes, including shape distortion, object merging, and identity inconsistency.
  • G. License: The artifact taxonomy is hierarchical across L1–L2–L3 categories, and Table 9 summarizes their distribution in ViF-CoT-4K.
  • G. License: The dataset uses forged samples from video-generation models whose conditioning modes include T2V, I2V, and TI2V, with closed-source parameter sizes marked N/A.
Loading 2512.15693v2…