🤖 AI

CoCoVideo: The High-Quality Commercial-Model-Based Contrastive Benchmark for AI-Generated Video Detection

This paper introduces CoCoVideo-26K, a high-quality dataset featuring semantically aligned real-fake video pairs from 13 commercial generators, and proposes CoCoDetect, a novel detection framework combining contrastive learning with confidence-gated multimodal reasoning to achieve state-of-the-art performance in identifying high-fidelity AI-generated videos.

Original authors: Huidong Feng, Wentao Chen, Jie Chen, Xinqi Cai, Ruolong Ma, Yinglin Zheng, Yuxin Lin, Ming Zeng

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Huidong Feng, Wentao Chen, Jie Chen, Xinqi Cai, Ruolong Ma, Yinglin Zheng, Yuxin Lin, Ming Zeng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the internet is a giant library. For a long time, the "fake books" in this library were easy to spot because they were printed on cheap, crinkly paper with blurry pictures. Detecting them was like spotting a typo in a handwritten note.

But recently, a new generation of "fake books" has arrived. These are printed on high-quality, glossy paper with pictures so sharp and stories so logical that even a human librarian might get fooled. This is the world of AI-generated video (AIGC). The problem is that the old "typo detectors" (existing AI tools) were trained on those cheap, crinkly papers. They don't know how to spot the subtle flaws in the new, high-quality fakes.

Here is what this paper does to fix that, explained in simple terms:

1. The New "Training Ground": CoCoVideo

The authors realized they couldn't teach their new detector using the old, low-quality fake videos. They needed a gym with high-quality equipment.

The Problem with Old Data: Previous datasets used open-source AI models that made videos looking a bit "off" (like a cartoon character blinking weirdly). Real-world commercial AI (the kind used by big companies) makes videos that look almost perfect.
The Solution (CoCoVideo-26K): The team built a massive new dataset called CoCoVideo.
- The "Twin" Strategy: Imagine you have a real photo of a forest. You ask 13 different high-end AI artists to paint a fake version of that exact same forest, starting from the exact same first frame.
- The Result: You now have "Real vs. Fake" pairs that are identical in story and starting point, but one is real and one is AI. This forces the detector to look for the tiny differences in texture and physics, rather than just guessing based on the story.

2. The New Detective: CoCoDetect

Once they had this high-quality training ground, they built a new detective system called CoCoDetect. Think of this detective as a two-person team working together:

Team Member A: The "Texture Scout" (Contrastive Learning)

What it does: This part of the system is like a forensic expert looking at the grain of the paper. It uses a standard video AI (R3D-18) to scan the video for tiny, invisible glitches—like a shadow that doesn't move right or a skin texture that looks too smooth.
How it learns: Because it was trained on those "Twin" pairs from CoCoVideo, it learned to spot the specific "fingerprint" of commercial AI, not just the obvious errors of old AI.

Team Member B: The "Logic Judge" (MLLM)

What it does: Sometimes, the video looks perfect to the Texture Scout. Maybe the shadows are right, and the skin looks real. But does the story make sense?
The "Confidence Gate": This is the clever part. The Texture Scout gives a "confidence score."
- High Confidence: If the Scout is 95% sure it's fake (or real), it makes the call immediately. Fast and efficient.
- Low Confidence (The "Maybe" Zone): If the Scout is unsure (e.g., "I think it's fake, but I'm only 60% sure"), it doesn't guess. Instead, it passes the case to the Logic Judge.
The Logic Judge's Job: This is a powerful AI that understands language and physics. It looks at the video and asks: "Wait a minute. In the video, a bird is flying backward while the wind is blowing forward. That's physically impossible. This must be fake."

3. How They Work Together

The system works like a security checkpoint:

Scan: The Texture Scout checks the video.
Gate: If the video is clearly fake or clearly real, the gate opens, and the decision is made.
Reasoning: If the video is tricky and the Scout is confused, the gate routes it to the Logic Judge. The Judge reads the scene, checks the laws of physics, and gives a final verdict.
Fusion: The final decision combines the Scout's "texture gut feeling" with the Judge's "logical reasoning."

Why This Matters (According to the Paper)

The authors tested their new detective against the old ones using their new high-quality dataset.

The Result: The old detectors (trained on low-quality fakes) failed miserably when faced with the new commercial-quality videos. They were like trying to find a needle in a haystack using a magnet that only works on iron, not steel.
The Winner: CoCoDetect, with its "Scout + Judge" team, performed significantly better. It proved that to catch high-quality fakes, you need a system that looks at both the tiny details (texture) and the big picture (logic/physics).

In short: The paper built a better "training school" using high-end commercial AI fakes and created a detective that uses a "gut check" for obvious fakes and a "logic check" for the tricky ones, making it much harder for high-quality AI videos to fool us.

Technical Summary: CoCoVideo and CoCoDetect

1. Problem Statement

The rapid advancement of Artificial Intelligence Generated Content (AIGC) has led to increasingly sophisticated video forgery, posing significant challenges to public discourse and societal security. While deepfake detection has progressed, current methods face two critical limitations:

Dataset Quality Gap: Existing benchmarks (e.g., GenBuster, GenVidBench) rely heavily on open-source video generation models. These models produce outputs with lower visual fidelity and weaker motion/scene consistency compared to state-of-the-art commercial systems. Consequently, models trained on such data tend to overfit low-quality generative patterns and fail to generalize to high-fidelity commercial AIGC videos.
Methodological Limitations: Traditional detection methods often focus on specific generative architectures or low-level manipulation cues (e.g., visual artifacts, frequency anomalies). They struggle to leverage high-level semantic cues, such as physical logic and scene coherence, which are crucial for distinguishing high-quality forgeries that lack obvious low-level artifacts.

2. Methodology

The paper introduces a two-pronged solution: a new high-quality dataset (CoCoVideo-26K) and a novel detection framework (CoCoDetect).

2.1. CoCoVideo-26K Dataset

CoCoVideo-26K is a contrastive, commercial-model-based benchmark designed to bridge the quality gap between research data and real-world threats.

Data Composition: The dataset comprises 26,000 video clips organized into semantically aligned "real–fake" pairs.
- Real Videos: Sourced from OpenVid-1M, filtered for high-fidelity, real-world scenes (~13,000 clips).
- Fake Videos: Generated using 13 distinct, state-of-the-art commercial video generation models (e.g., Sora, Kling, Runway Gen4, Jimeng 3.0, Midjourney v7). Each model contributes 1,000 videos.
Contrastive Structure: Each real video is paired with a synthetic counterpart generated from the same first frame and identical text prompt. This ensures strict semantic alignment, allowing for fine-grained analysis of differences in texture, motion, and scene consistency under matched content conditions.
Scale and Diversity: The dataset covers diverse categories including human interactions, food preparation, plant growth, and natural landscapes, with a consistent duration of ~5 seconds and frame rates of 24–30 FPS.

2.2. CoCoDetect Framework

CoCoDetect is a detection framework that integrates contrastive learning with a confidence-gated Multimodal Large Language Model (MLLM) reasoning mechanism.

Backbone: The system utilizes an R3D-18 network to extract spatio-temporal representations from input video batches.
Dual-Head Architecture:
1. Confidence Head: A three-layer MLP followed by a Sigmoid activation outputs a confidence score ( $p \in [0, 1]$ ) indicating the likelihood of a video being real.
2. Projection Head: A two-layer MLP maps features into a 128-dimensional embedding space for contrastive learning.
Loss Function:
- Confidence Loss: Binary Cross-Entropy (BCE) for authenticity classification.
- Paired Contrastive Loss: Unlike traditional supervised contrastive learning that clusters all samples of the same class, this loss specifically enforces separation between semantically aligned real–fake pairs (sharing the same first frame and prompt). It uses a hinge penalty to ensure the cosine similarity between a real-fake pair remains below a specific threshold, without constraining non-paired samples.
Confidence-Gated MLLM Inference:
- Routing Mechanism: During inference, if the confidence score $p$ falls within an uncertain range (defined by threshold $\tau$ ), the sample is routed to an MLLM (LLaVA-NeXT-Video-7B). High-confidence samples are classified directly by the backbone.
- MLLM Reasoning: For uncertain cases, the MLLM analyzes key frames and the confidence score to reason about physical plausibility and scene coherence. It outputs a structured JSON containing a binary judgment, a semantic confidence score, and a natural language explanation.
- Decision Fusion: The MLLM's output is mapped to the same semantic space as the backbone's confidence and fused via an adaptive weighting strategy to produce the final classification.

3. Key Contributions

CoCoVideo-26K Benchmark: The establishment of a high-quality dataset featuring contrastive real–fake pairs from 13 commercial models, providing a new standard for evaluating AIGC detection under realistic, high-fidelity conditions.
CoCoDetect Framework: A novel architecture combining dual-head contrastive learning (for texture-level discrimination) with a confidence-gated MLLM mechanism (for semantic-level reasoning), enabling systematic judgment across both feature levels.
Empirical Validation: Comprehensive experiments demonstrating state-of-the-art performance on the CoCoVideo test set and strong generalization capabilities across multiple public open-source benchmarks.

4. Experimental Results

4.1. Performance on CoCoVideo-26K

On the proposed test set, CoCoDetect achieved the best average performance across Accuracy (90.69%), F1-score (90.62%), Recall (89.95%), and AUC (95.93%), outperforming strong baselines like DeMamba, VideoMAE, and 3D ResNeXt.

Observation: The method showed a conservative decision strategy, prioritizing precision over recall in some models, effectively reducing false alarms by selectively engaging the MLLM for uncertain cases.
Limitation: Performance degraded slightly on models like Jimeng and Veo3 where resolution inconsistencies between real and fake pairs disrupted semantic alignment.

4.2. Cross-Dataset Generalization

Evaluated on five public benchmarks (GVD, GVF, GenVideo, GenVidBench, GenBuster), CoCoDetect demonstrated robust generalization:

It achieved the highest accuracy on four of the five benchmarks.
On GenBuster, which uses older versions of the same commercial models, CoCoDetect maintained strong performance, suggesting it relies less on specific generation artifacts and more on generalizable semantic and texture cues.
On GenVideo, performance dropped in F1-score due to the dataset's prevalence of short, nearly static clips, which disrupted the temporal coherence required for the MLLM's motion analysis.

4.3. Ablation Studies

Ablation experiments confirmed the necessity of both components:

Removing the Projection Head (contrastive learning) reduced AUC from 95.93% to 91.39%.
Removing the MLLM (relying only on confidence scores) reduced AUC to 95.46%.
The full model achieved the highest scores, validating that low-level contrastive features and high-level semantic reasoning work synergistically.

4.4. Case Studies

High Confidence: The contrastive network correctly identified obvious texture artifacts without MLLM intervention.
Low Confidence (Rescued): In ambiguous cases, the MLLM successfully identified semantic errors (e.g., a suspension bridge misinterpreted as a chimney) that the backbone missed, correcting the prediction.
Failure Case: In simple natural scenes with subtle artifacts and no obvious semantic inconsistencies, both modules failed, highlighting the current difficulty in detecting forgeries in low-complexity environments.

5. Significance and Claims

The paper claims that CoCoVideo-26K addresses the critical gap between research datasets and real-world commercial AIGC threats, providing a more reliable benchmark for training and evaluation. CoCoDetect is presented as a robust framework that overcomes the limitations of purely texture-based detection by integrating semantic reasoning for uncertain cases.

The authors emphasize that their approach validates the effectiveness of combining fine-grained contrastive learning with high-level semantic verification. They conclude that while the method demonstrates state-of-the-art performance and strong generalization, future work will focus on expanding the dataset with audio modalities and richer semantic annotations to further enhance trustworthy AIGC content detection.