Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning

AI summary

Overview

Research area: Computer vision / multimodal large language models, specifically explainable detection of AI-generated video.

Technical level: Advanced. The paper combines a purpose-built human-annotated dataset, full-parameter supervised fine-tuning of a 7B vision-language model, and GRPO-based reinforcement learning with a custom reward design.

Scope: The paper introduces Skyra, a specialized multimodal large language model (MLLM) that detects AI-generated videos by locating human-perceivable visual artifacts and using them as grounded, spatio-temporally localized evidence, together with two supporting resources: the ViF-CoT-4K training dataset and the ViF-Bench evaluation benchmark.

What This Paper Is About

Most existing AI-generated video detectors output only a binary "real" or "fake" label, giving a human no way to see why the judgment was made. Meanwhile, general-purpose MLLMs — even strong ones — perform near randomly on this task and tend to latch onto superficial cues like lighting or visual quality rather than the physics-violating artifacts humans actually notice.

Skyra's goal is to close that gap: a model that first finds specific artifacts in space and time (e.g., an object disappearing at a given timestamp inside a given bounding box), then grounds its real/fake decision and its written explanation on those findings.

Key Contributions

1. Skyra, a specialized MLLM for AI-generated video detection via grounded artifact reasoning. It identifies human-perceivable artifacts and uses them as spatio-temporally grounded evidence for both classification and explanation.

2. ViF-CoT-4K, described as the first large-scale AI-generated video artifact dataset with fine-grained human annotations, including artifact type, textual explanation, timestamps, bounding boxes, and chain-of-thought (CoT) explanations.

3. A two-stage training strategy: a cold-start supervised fine-tuning (SFT) stage that gives the base model basic artifact perception (producing Skyra-SFT), followed by a reinforcement learning stage using GRPO with redesigned rewards (producing Skyra-RL).

4. ViF-Bench, a benchmark of 3K high-quality samples generated by more than ten state-of-the-art video generators, with real and fake samples aligned in semantics and format to reduce shortcut signals.

Main Findings

• Skyra leads on ViF-Bench. Skyra-RL(7B) reaches a mean accuracy of 91.02%, recall of 88.35%, and F1 of 90.27%, versus 90.11% / 84.65% / 88.76% for Skyra-SFT(7B). The paper reports this as +26.73% absolute accuracy and +17.27% F1 over the second-best method, DeMamba (mean accuracy 64.29%, recall 96.66%, F1 73.00%).

• Gains over MLLM baselines are large. The paper reports +34.12% accuracy, +24.57% recall, and +32% F1 relative to MLLM baselines. Off-the-shelf MLLMs perform poorly: Video-LLaMa-3(7B) mean accuracy 50.91%, Qwen2.5-VL(3B) 50.7%, Qwen2.5-VL(7B) 50.57%, Qwen2.5-VL(72B) 51.26%, InternVL-3(8B) 48.58%, GPT-4.1-mini 54.08%, Gemini-2.5-flash 53.36%, and the MLLM-based detector BusterX++(7B) 56.90%.

• RL improves over SFT, especially on I2V samples. The paper reports a +3.74% recall gain from the RL stage on the harder image-to-video (I2V) cases.

• Out-of-domain performance on GenVideo. Skyra-RL(7B) achieves 71.78% average accuracy, 45.60% recall, and 59.00% F1 on GenVideo, reported as +11.07% accuracy over the best binary detector and +7.8% accuracy / +16.9% recall over Skyra-SFT. By running RL for one epoch on only 2.2K samples randomly selected from the GenVideo-100K training set, with no additional human annotation, the resulting Skyra-RL-GenVideo(7B) reaches 91.00% average accuracy, 87.66% recall, and 90.00% F1 — described as +19.22% accuracy, +42.06% recall, and +31% F1 over Skyra-RL.

• Robustness under degradations. Skyra-RL(7B) starts at 91.02% / 88.35% / 90.27% (accuracy / recall / F1) on unperturbed data and remains state-of-the-art under all five tested degradations: compression (80.80 / 88.64 / 81.93), transformation (83.26 / 96.37 / 85.17), Gaussian noise (83.48 / 96.34 / 85.33), light transform (83.26–90.66 / 85.78–96.37 / 85.17–89.62 for − / +), and color transform (83.26–90.67 / 85.81–96.37 / 85.17–89.64).

• Chain-of-thought is essential. Removing CoT drops SFT accuracy from 90.11% to 54.04% (−36.07) and F1 from 88.76% to 16.72% (−72.04).

• Cold-start initialization is essential. Removing it leaves the model at 50.09% accuracy (−40.93), 0.18% recall (−88.17), and 0.37% F1 (−89.90) — worse even than training without CoT. The RL stage alone cannot equip the base model with artifact-identification ability.

• Reward design matters. Replacing the asymmetric accuracy reward with a symmetric one causes overfitting to "Fake": accuracy falls 14.78 points to 76.24% and F1 falls 9.62 points to 80.65%, while recall rises 10.72 points to 99.07%.

• The RL stage adds a further gain. Removing it leaves the SFT numbers unchanged (90.11 / 84.65 / 88.76 versus the RL model's 91.02 / 88.35 / 90.27).

• The artifact taxonomy is hierarchical. Layer 1 splits into Low-level Forgery (perceptual quality artifacts) and Violation of Laws (physical and logical inconsistencies). Layer 2 has eight categories: Color/Light Anomaly, Texture Anomaly, Motion Forgery, Object Inconsistency, Interaction Inconsistency, Violation of Causality, Violation of Commonsense, and Unnatural Movement. Layer 3 gives the most fine-grained observable artifacts — for example, Object Inconsistency divides into Abnormal Object Disappearance, Abnormal Object Appearance, and Person Identity Inconsistency.

Methodology in Plain English

The authors start from an observation about how people spot fake videos: humans take in the overall scene, then actively hunt for things that break physical or temporal coherence — an object vanishing, a motion that does not make sense. Skyra is built to imitate that, treating such cues as intrinsic evidence that is model-agnostic and not tied to any specific generator's fingerprint.

Building the data. They gathered roughly 3.5K real videos from Panda-70M and 1.5K from Kinetics-400, plus high-resolution videos from HD-VILA-100M for the benchmark. Several MLLMs generated descriptions of the real videos, which after manual inspection became prompts for text-to-video generators; for image-to-video, the first frame of a real video served as the condition. GPT-4o-mini ran an automatic filtering step to keep semantics consistent between real and generated pairs. Training used Wan2.2-TI2V-5B, Wan2.1-T/I2V-1.3B, CogVideoX-1.5-5B, and HunyuanVideo; evaluation used newer models including Wan2.2-T/I2V-A14B, LTX-Video-13B, MiniMax-Hailuo, and Sora-2.

Annotating it. Professional annotators identified every visible artifact, labeling its type, a textual explanation, timestamps, and bounding boxes. Crucially, they viewed each generated video next to its real counterpart and had to point to matching real evidence for each fake evidence, which validates that the artifacts are generation-induced rather than compression degradation. Gemini-2.5-Pro then turned these labels into step-by-step chain-of-thought explanations using self-curation (an observe-understand-draft-review-conclude process) and in-context learning with definitions and examples per artifact type.

Training in two stages. Stage one fine-tunes Qwen2.5-VL-7B on ViF-CoT-4K with a standard cross-entropy loss over a fixed response template: an outer <thinking>...</thinking> block plus an <answer>Fake/Real</answer> block, with forgery type, time range, and bounding box tags for fake videos, and matching temporal-spatial tags for real ones. The model samples 16 frames per video at 256p, with full-parameter fine-tuning at batch size 1 per device, 5 epochs, and a learning rate of 1e-5 on 8 NVIDIA H200 GPUs.

Stage two uses GRPO reinforcement learning, motivated by the observation that human annotators themselves struggled on high-quality samples while low-quality samples carried noisy labels. The reward is a weighted sum of an accuracy reward (weight 0.8) and a check reward (weight 0.2). The accuracy reward is deliberately asymmetric: +1.0 for a correct prediction, 0.0 for calling a fake video real, and −0.2 for calling a real video fake — because finding one artifact suffices to prove fakeness, while proving realness requires ruling out all inconsistencies. The check reward is min(ln(1 + N_check), ln(1 + 3)), rewarding the model for producing multiple valid grounded inspection blocks, capped at three. RL uses an actor learning rate of 5e-7 and a KL coefficient of 0.02.

Why This Matters

Impact on research. The paper reframes AI-generated video detection from opaque binary classification to grounded, evidence-based reasoning, and argues that binary detectors are stuck in a continuous adversarial cycle where previously discriminative features go stale as generators improve. It contributes a human-annotated artifact dataset, a benchmark with real/fake pairs matched in semantics and format to suppress shortcut learning, and an open, reproducible recipe (code, models, and datasets are stated to be publicly available) including a demonstration that the RL stage can adapt a model to a new domain in one epoch on 2.2K samples with no new human annotation. It also documents a negative result that is useful to the field: pure RL without cold-start initialization fails on this task, unlike the pattern seen in DeepSeek-R1-Zero-style training.

Real-world applications:

• Content moderation and platform trust-and-safety workflows, where the misuse of AI-driven video generation is the paper's stated motivation.
• Media forensics and journalism verification, where an analyst needs to see where in the frame and when in the timeline the evidence lies, not just a score.
• Human-in-the-loop review pipelines, since the paper explicitly argues that interpretable detection supports scenarios requiring manual verification.
• Detection under degraded real-world conditions — compression, zoom, noise, and light or color transforms — where the robustness study shows the model holds up.

Industry relevance. The commercial video generation landscape the paper enumerates (Sora-2, Kling, MiniMax-Hailuo, Pika, Gen-4-Turbo, PixVerse, SkyReels, LTX-Video) is exactly the ecosystem that platforms, newsrooms, and regulators must contend with. A 7B open model that runs on a defined frame budget and produces a written, localized rationale is a practical candidate for deployment, subject to the paper's own caveats about calibration and human oversight.

Future Directions

• Broaden the media distribution. The authors state that ViF-CoT-4K and ViF-Bench are bound by the specific generators and collection pipeline used, and that the benchmark does not yet cover emerging distributions such as ultra-long videos or non-photorealistic, stylized content.

• Assess intent, context, and societal harm. Skyra does not evaluate why a video was made or what harm it might cause; the paper flags this as outside the current scope. It also warns that its natural-language rationales, while crafted to be persuasive and readable, may be overconfident or partially hallucinated, which points to calibrated uncertainty estimation as a needed extension.

• Reduce reliance on manual annotation. The paper positions RL as a way to improve adaptability to new domains and mitigate the need for iterative manual annotation, demonstrated by the one-epoch GenVideo adaptation. Extending that to new generators and artifact types without fresh labels is an open direction.

• Improve temporal reasoning. The paper finds that current methods, including prior MLLM-based detectors, struggle with complex temporal dynamics and fine-grained reasoning — the RL stage's largest gain over SFT is reported on I2V samples, suggesting this remains a frontier.

Target Audience

Researchers and engineers working on AI-generated content detection, deepfake and synthetic media forensics, and multimodal large language models — particularly those interested in explainable detection, reinforcement learning post-training for vision-language tasks, or curating fine-grained human-annotated datasets. It is also relevant to trust-and-safety practitioners and policy-oriented readers who need to understand both the capability and the current limits of automated video authenticity tools. Readers should be comfortable with MLLM training terminology (SFT, GRPO, reward shaping, chain-of-thought supervision); the artifact taxonomy and results sections are accessible to a broader audience.

Authors’ abstract

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.