Skyra: Explainable Video Forgery Detection

• Skyra is a specialized multimodal LLM that detects AI-generated videos by analyzing spatio-temporal inconsistencies and localizing visual artifacts.
• It leverages the large-scale ViF-CoT-4K dataset and employs a two-stage training strategy—supervised fine-tuning followed by reinforcement learning—to achieve state-of-the-art accuracy.
• The model outputs human-interpretable explanations, detailing artifact types, temporal spans, and spatial bounding boxes to facilitate rigorous human-in-the-loop auditing.

Skyra is a specialized multimodal LLM (MLLM) designed to detect AI-generated videos by identifying and localizing human-perceivable visual artifacts, providing not only precise detection but also human-interpretable, grounded explanatory evidence. Unlike binary classifiers that output only a “Real” or “Fake” verdict, Skyra emulates human reasoning by analyzing spatio-temporal inconsistencies, generating a chain-of-thought (CoT) that details visual artifacts linked to synthetic content. To support this capability, Skyra leverages ViF-CoT-4K, a large-scale, fine-grained, human-annotated dataset of video artifacts, and achieves robust performance through a two-stage supervised and reinforcement learning strategy. Evaluations demonstrate state-of-the-art accuracy and explanation fidelity across diverse, modern AI video generators and robustness scenarios (Li et al., 17 Dec 2025).

1. Conceptual Foundations and Objectives

Skyra addresses escalating concerns regarding the misuse of AI-generated video by advancing both explainability and detection reliability. Its core objectives are:

• Accurate classification of videos as “Real” or “Fake,” surpassing standard binary detectors.
• Generation of explicit, human-interpretable explanations, including annotation of detected artifacts with category, spatio-temporal localization (bounding boxes and time spans), and natural language justification.
• Facilitation of human-in-the-loop auditing by providing chain-of-thought reasoning and artifact evidence, enhancing model transparency.

Skyra operationalizes these objectives by modeling the human strategy of interrogating videos for spatio-temporal inconsistencies, outputting a structured protocol for both the detection process and its justifications.

2. Model Architecture and Output Structure

Skyra extends the Qwen2.5-VL-7B-Instruct MLLM as its base model with the following architectural components:

• Visual Encoder: Uniformly samples $T=16$ frames at $256p$ resolution per video, feeding them into a pretrained vision transformer to produce per-frame feature maps.
• Text Encoder: Tokenizes all input prompts, including interleaved “[T= ] <image>” markers for temporal correspondence, using a transformer architecture.
• Fusion and Decoder: Cross-modal cross-attention layers fuse visual and textual embeddings. An auto-regressive language decoder then outputs:
  • A <thinking> block: chain-of-thought reasoning detailing artifact identification and justification.
  • A <answer> token: either “Fake” or “Real.”
  • For “Fake”: Auxiliary tags including <type>…</type>, <t>[t_start,t_end]</t>, and <bbox>[x_min,y_min,x_max,y_max]</bbox> for artifact type and localization.
  • For “Real”: The same tag format, used as structured “inspection evidence” indicating systematic artifact search yielded no findings.

This output format enables both automated and human review of detections and rationales.

3. ViF-CoT-4K: Dataset and Annotation Protocol

The ViF-CoT-4K dataset is central to Skyra’s methodology, comprising:

• Data Composition: Over 4,000 AI-generated videos (from generators such as Wan 2.1, Wan 2.2, CogVideoX, HunyuanVideo) paired with a roughly equal number of real videos derived from Panda-70M, Kinetics-400, and HD-VILA-100M.
• Artifact Taxonomy:
  • Level 1: Low-Level Forgery (17.2%), Violation of Laws (82.8%)
  • Level 2: Eight subcategories (e.g., Texture Anomaly, Object Inconsistency, Violation of Commonsense)
  • Level 3: Eighteen fine-grained artifact types (e.g., “Structure Anomaly,” “Unnatural Human Movement,” “Text Distortion”)
• Annotation Schema: Each fake video includes (1) artifact type (L3), (2) textual explanation, (3) temporal span $[t_{\mathrm{start}}, t_{\mathrm{end}}]$, (4) spatial bounding box $(x_{\min}, y_{\min}, x_{\max}, y_{\max})$, and (5) a paired annotation on the corresponding real video.
• Chain-of-Thought Expansion: Concise human labels are expanded into multi-step reasoning scripts using Gemini-2.5-Pro prompts, adhering to a standardized observe–understand–draft–review–conclude protocol in JSON format.

A data analysis of ViF-CoT-4K shows dominant artifact types: Object Inconsistency (28.1%), Violation of Commonsense (27.8%), Texture Anomaly (11.2%), among others.

4. Training Strategy: Supervised Fine-Tuning and Reinforcement Learning

Skyra training proceeds in two distinct stages:

4.1. Supervised Fine-Tuning (Skyra-SFT)

• Initialization: From Qwen2.5-VL-7B-Instruct, using ViF-CoT-4K for targets.
• Inputs: Frames $v$ plus prompt $t$ (system and user chain-of-thought template).
• Targets: Full token sequence $y^*$ including CoT reasoning, <answer>, and artifact localization tags.
• Loss Function: Cross-entropy loss over all output tokens,

  $$\mathcal{L}_{\mathrm{SFT}} = -\sum_{t=1}^T \log p_\theta(y^*_t \mid y^*_{<t}, t, v)$$

• Optimization:
  • Learning rate $= 1 \times 10^{-5}$
  • Batch size 1 per GPU, 5 epochs, 8×H200 GPUs.

4.2. Reinforcement Learning (Skyra-RL)

• Algorithm: Group Relative Policy Optimization (GRPO).
• Reward Function: Weighted sum of accuracy ($w_{\mathrm{acc}}=0.8$) and inspection ($256p$0) rewards:

  $256p$1

  • Asymmetric accuracy reward:

  $256p$2 - Inspection reward incentivizes correct structured outputs and at least one valid artifact check block.

• KL Regularization: Coefficient 0.02, actor learning rate $256p$3.

• Duration: Approximately one epoch (4,000 samples) for RL refinement.

Ablation studies reveal that both CoT supervision and the two-stage training are critical for high performance, with major accuracy and F1 drops if either is omitted.

5. Evaluation: Benchmarks, Metrics, and Results

Skyra's efficacy is validated through extensive benchmarking:

• Benchmarks:

  • ViF-Bench: 3,000 videos from 11 state-of-the-art video generators, matched in semantics and structure for real/fake.
  • GenVideo: 8,000 out-of-domain (OOD), lower-quality videos from older models.
  • Robustness: JPEG compression, geometric transforms, Gaussian noise, lighting/color jitter.
• Metrics: Accuracy (Acc), Recall on “Fake” class (R), F1-score.

Key Quantitative Results (ViF-Bench)

Method	Acc	R	F1
DeMamba	64.29	96.66	73.00
Qwen2.5-VL(7B)	50.57	3.53	6.63
BusterX++	56.90	14.40	21.94
Skyra-SFT	90.11	84.65	88.76
Skyra-RL	91.02	88.35	90.27

• Skyra-SFT outperforms the best binary detector by +25.8% accuracy and +18.7% F1-score.
• Skyra-RL further improves accuracy (+0.9%) and F1 (+1.5%), primarily by enhancing recall on difficult I2V samples.
• On OOD GenVideo (without extra labels), Skyra-RL achieves 57.21% accuracy (+11.07% over DeMamba). After one epoch of RL using 2,200 GenVideo samples, Skyra-RL-GenVideo yields 91.00% accuracy, 87.66% recall, and 90.00% F1.

Robustness and Qualitative Examples

Skyra-RL maintains >80% accuracy and F1 under all tested data corruptions. Qualitative outputs include precise artifact descriptions (e.g., “Structure Anomaly at T=[1.5–2.0]s, bbox=[…]: twisted fence lines flickering”) and structured negative findings for real videos.

6. Strengths, Limitations, and Future Directions

Strengths:

• Detection accuracy exceeding 90% on in-distribution data with strong OOD generalization.
• Fine-grained, grounded explanations coupling artifact type with spatio-temporal evidence.
• Modular transferability: the combination of cold-start SFT and RL supports rapid adaptation to new domains using minimal incremental supervision.

Limitations:

• Coverage is restricted to 16-frame clips and contemporary T2V/I2V model artifacts.
• Model explanations may hallucinate or cite non-causal features, indicating imperfect fidelity in artifact localization.
• Skyra does not address intent, broader contextual analysis, or assessment of societal harm, focusing instead on evidence-driven auditing workflows.

Future research aims:

• Expansion to ultra-long videos (>1 minute) and non-photorealistic or stylized generative models.
• Incorporation of uncertainty estimation and mechanisms for calibrated active human feedback.
• Distributed or multitask architectures integrating detection with watermarking or provenance tracking.
• Exploration of hierarchical spatio-temporal attention mechanisms to further increase the localization resolution of detected artifacts.

Skyra exemplifies a new paradigm in explainable video forgery detection, combining high-precision verdicts with structured, inspectable chains of reasoning (Li et al., 17 Dec 2025).