ViF-CoT-4K: AI Video Artifact Dataset

• ViF-CoT-4K is a large-scale, fine-grained dataset that pairs AI-generated videos with real counterparts for detecting and explaining video artifacts.
• It employs a three-level hierarchical annotation scheme to localize artifacts with precise temporal spans and spatial bounding boxes.
• The dataset supports supervised fine-tuning with chain-of-thought rationales, advancing robust AI video forgery detection models.

ViF-CoT-4K is a large-scale, fine-grained, human-annotated dataset specifically constructed for the detection and explanation of AI-generated video artifacts. It is designed to support supervised fine-tuning of multimodal LLMs (MLLMs) in the context of explainable AI-generated video detection. Each of the ≈5,467 AI-generated videos in ViF-CoT-4K is paired with a semantically aligned, duration-matched real video, providing a unique and rigorous resource for spatio-temporal artifact detection, localization, and explanation tasks (Li et al., 17 Dec 2025).

1. Dataset Scope and Composition

ViF-CoT-4K (Video Inference with Chain-of-Thought, 4,000+) constitutes the first large-scale artifact dataset for AI-generated video detection to feature fine-grained, human-perceivable annotations with both precise temporal and spatial localization. The dataset covers AI-generated videos sampled from multiple generative systems, including both open-source and commercial text-to-video (T2V) and image-to-video (I2V) models:

Generator	Sample Count
Wan2.1-T2V-1.3B	750
CogVideoX-1.5-T2V	744
CogVideoX-1.5-I2V	760
Wan2.2-TI2V-5B (T2V)	747
Wan2.2-TI2V-5B (I2V)	748
HunyuanVideo (T2V)	750
HunyuanVideo (I2V)	968
Total (with real counterparts)	≈5,467

Videos are stored at up to 540p resolution, with 16 frames per video uniformly sampled and resized to 256p for model training. The typical duration of each real–fake video pair is 5–15 seconds, with lengths matched to avoid spurious cues. The real videos originate from Panda-70M (3.5K samples) and Kinetics-400 (1.5K samples), ensuring a diversity of content. Prompts used for video generation are systematically extracted from real video descriptions by MLLMs and filtered for semantic consistency via GPT-4o-mini (Li et al., 17 Dec 2025).

2. Hierarchical Annotation Schema

The annotation protocol employs a three-level hierarchical taxonomy, organized as follows:

• L1 (Top):
  • Low-Level Forgery (17.2% of annotations)
  • Violation of Laws (82.8% of annotations)
• L2 (Intermediate):
  • Under Low-Level Forgery: Texture Anomaly, Color/Lighting Anomaly, Motion Forgery
  • Under Violation of Laws: Object Inconsistency, Interaction Inconsistency, Unnatural Movement, Violation of Causality, Violation of Commonsense
• L3 (Fine-grained): 22 specific artifact types, including Structure Anomaly, Texture Jittering, Unnatural Blur, Color Over-Saturation, Lighting Inconsistency, Camera Motion Inconsistency, Abnormal Object Disappearance/Appearance, Person Identity Inconsistency, Shape Distortion, Abnormal Interaction patterns, Unnatural Human/Animal/Object Movement, Violation of Physical Law, and Text Distortion.

A statistical breakdown is:

Level	Category	% of Annotations
L1	Low-Level Forgery	17.2
L1	Violation of Laws	82.8
L2 (ex.)	Object Inconsistency	28.1
L2 (ex.)	Violation of Commonsense	27.8
L3 (ex.)	Shape Distortion	15.2
L3 (ex.)	Text Distortion	14.1
L3 (ex.)	Abnormal Body Structure	10.5

Full annotation count ratios for each layer (L1–L3) are enumerated in (Li et al., 17 Dec 2025) Table 1.

3. Data Collection and Annotation Process

Generation and annotation follow a highly controlled pipeline:

• Pairing: Each AI-generated video is paired with a real video that is both semantically and temporally aligned. Duration is matched using automated filtering by GPT-4o-mini; content prompts are extracted via MLLMs.
• Annotation Interface: A side-by-side annotation GUI is used by professional domain experts. Each detected artifact is labeled with:
  1. L3 artifact type
  2. Textual explanation (for both fake and real videos)
  3. Temporal span $[t_\mathrm{start}, t_\mathrm{end}]$ in seconds
  4. Spatial bounding box $$[x_\mathrm{min}, y_\mathrm{min}, x_\mathrm{max}, y_\mathrm{max}]$$

• Normalization: In real videos, corresponding “normal” regions are annotated to disambiguate genuine content from generative defects or compression/motion blur artifacts.

• CoT Rationales: Concise human annotations are expanded automatically to full reasoning chains (observe–understand–draft–review–conclude) via Gemini-2.5-Pro, yielding trainable chain-of-thought (CoT) explanations.
• Quality Assurance: The protocol includes multi-pass review, inter-annotator consistency checks, and self-curation.

The result is frame-level bounding box and sub-second temporal granularity for artifact localization.

4. Statistical Properties and Artifact Distribution

Artifact distributions reveal that most detected errors (82.8%) fall under the “Violation of Laws” L1 category, dominated by object inconsistency, commonsense violations, and structural distortions rather than low-level visual forgeries. Specific breakdowns (L2/L3) include:

L2/L3 Category	% of Annotations
Object Inconsistency (L2)	28.1
Shape Distortion (L3)	15.2
Abnormal Body Structure (L3)	10.5
Text Distortion (L3)	14.1
Unnatural Human Movement (L3)	6.6
Violation of Physical Law (L3)	4.1

Temporal annotation varies from 0.5 s up to full-clip length, and bounding boxes are defined at the per-frame level.

5. Use in Supervised Fine-Tuning Workflows

ViF-CoT-4K was constructed as the primary supervised fine-tuning (SFT) corpus for Skyra—a multimodal LLM based on Qwen2.5-VL-7B-Instruct. The SFT workflow involves:

• Input: Chain-of-thought text prompt $t$ and video frame batch $v$ (16 frames per video, 256p).
• Output: Autoregressive generation of three blocks:
  1. <thinking>: stepwise reasoning
  2. <answer>: Fake/Real label
  3. Per-artifact or per-region tags (<type>, <t>, <bbox>)

• Loss Function: Standard cross-entropy over autoregressive token outputs:

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

• Optimization: AdamW, batch size 1 per GPU, 5 epochs, $1 \times 10^{-5}$ learning rate, on 8× NVIDIA H200 GPUs.
• Ablation: The SFT stage on ViF-CoT-4K is necessary for model “cold-start” performance prior to any reinforcement learning stage.

The prompt structure is enforced via:

<thinking>[…chain‐of‐thought…]</thinking> <answer>[Fake/Real]</answer>

with further spatial-temporal tags as output.

6. Impact and Significance

ViF-CoT-4K establishes a new standard for the supervised detection and interpretation of AI-generated video artifacts:

• Granular explanations: Enables models not only to output binary classification but also to localize, categorize, and provide human-interpretable explanations for each detected artifact.
• Benchmark resource: Underpins rigorous evaluation and ablation for new video forgery detectors, especially MLLMs leveraging chain-of-thought rationale and spatial-temporal evidence.
• Coverage and diversity: Spans open-source and proprietary generation models, offering broad coverage of artifact diversity present in contemporary generative models.

A plausible implication is that ViF-CoT-4K’s explicit coupling of artifacts, temporal localization, spatial tagging, and natural-language reasoning is essential for the next generation of explainable, reliable AI-generated media detection frameworks (Li et al., 17 Dec 2025).