Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableNeurohive→ original

Tencent open-sources an agent that detects generated video with 95.2% accuracy

Tencent has open-sourced WeVisForensics, an AI agent that checks a video like a team of six narrow specialists. They look at skin and shadows, motion physics, common sense, on-screen text, voice, and speech tempo. Across four benchmarks it correctly classifies 95.2% of videos on average, 9 points better than the previous leader, VideoVeritas. Every verdict comes with timecodes, so a moderator can check the decision frame by frame. The weights are open, and a single 24 GB GPU is enough to run it.

AI-processed from Neurohive; edited by Hamidun News
Tencent open-sources an agent that detects generated video with 95.2% accuracy
Source: Neurohive. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceNeurohive ↗
Editor-in-Chief
Key metric #1On the new AgenticCue benchmark (Sora 2, Veo 3.1, Kling 3.0), accuracy is 97–98.8%
Key metric #2Joint training reached 95.2%, versus 93.2% for a trained agent alone
Key metric #3On MintVid, it catches Sora 2 fakes in 64.6% of cases and Jimeng fakes in 67%

Generated videos have become so smooth that fewer and fewer people can tell them apart by eye, while platforms and newsrooms need to decide quickly which clip goes into the feed. Researchers at WeChat Vision (Tencent) and Nanjing University have released WeVisForensics, an open AI agent built on Qwen3.5-4B that checks video like a team of six specialists.

Across four test sets (benchmarks), it correctly classifies 95.2% of clips on average, 9 percentage points more than the previous leader, VideoVeritas. Say a moderator receives a clip of a cat behind the wheel with a crooked sign in the background: the agent will point to the seconds where the letters fuse together and explain why the scene is unrealistic.

This is worth reading for anyone building moderation or content verification: the code and weights are open, and everything runs on a single GPU with 24 GB.

How the expert team works

Imagine an editor who receives a suspicious clip and hands it out to colleagues. One examines skin, hair, fingers, teeth, light, and shadows. A second watches the physics: how bodies move, how liquid flows, what happens when an object passes behind an obstacle. A third judges common sense, such as a cat holding a steering wheel the way a human would. A fourth reads the text. Two more tools listen to the voice and the speech tempo.

That is how the system works. The main agent first skims up to 16 frames sampled evenly from the video, and it knows the clip's duration and whether it has sound. Then, step by step, it decides which expert to call, which segment to show that expert, and when it has enough data for a verdict. It makes no more than four calls, and it cannot call the same expert twice.

The four visual experts run on the same Qwen3.5-4B, but each has its own LoRA adapter (a small fine-tuned add-on). PaddleOCR finds the text, and the expert then inspects enlarged crops for fused and broken letters. Two non-trainable audio tools estimate the speaker's gender and age and look for unnaturally even speech tempo. Each expert returns a verdict (artifact present, absent, or cannot be determined), a confidence level, and an explanation that references specific seconds, so the decision can be double-checked against the frames. The experts do not treat motion blur, compression, or stylization as signs of generation.

What is new in training

Earlier agentic detectors, such as ARGUS, trained only the main agent, while the experts stayed frozen. If an expert cannot see an artifact itself, no amount of routing will fix that. Updating both sides at every step is also a bad idea: the main agent decides what the experts will see, and their answers change its decisions, so each becomes a moving target for the other.

The authors decoupled the training speeds and called the method DTPO (Dual-Timescale Policy Optimization). The main agent is updated on every batch via GRPO, with a reward for a correct verdict without protocol violations. The experts are fine-tuned once every 20 steps on accumulated calls.

The teacher, a slowly updated copy of the expert, sees the frames, the label, and the artifact intervals, while the student sees only the frames and is pulled toward the teacher's answers. In addition, a frozen LLM analyzes the experts' mistakes and proposes rules such as "do not count hair smeared by fast motion as a defect." A rule is accepted only if the number of correct verdicts goes up and no previously correct verdict breaks.

Full training takes about a day.

The ablations show what each part contributes. Training only the experts raises average accuracy from 91.3% to 92.0%, training only the agent raises it to 93.2%, and training both raises it to 95.2%. For the physics expert, F1 on GenBuster-Bench rose from 66.5% to 97.8%, and the share of "cannot be determined" answers fell from 26.9% to 0.3%. If the experts are updated on every batch instead of every 20 steps, accuracy drops on all six MintVid generators.

The numbers, the limitations, and what to do about them

On GenBuster-Bench, the agent scores 97.7% versus 94.3% for VideoVeritas. On GenBuster-Bench++ it scores 97.5% versus 91.4%, and on the six MintVid generators 88.0% versus 82.2%. The binary detectors NSG-VD and ReStraV, trained on the same data, stayed at guessing level: 46.4% and 51.9% on GenBuster-Bench. Gemini-3.1-Pro finds 99% of fakes on GenBuster-Bench++, but it also flags real videos as fake, so its accuracy is 54.3%.

On the new AgenticCue benchmark, with clips from Sora 2, Veo 3.1, and Kling 3.0, the agent scores 97–98.8% in all five artifact categories, while VideoVeritas manages only 69.3% on audio. When the agent calls the physics, text, common-sense, or audio expert, an artifact of that type is actually present in the video in 94.3–99.7% of cases. Humans blindly compared 200 pairs of answers from the original and trained experts: the trained experts were better in 65 pairs and the originals in 20, and the share of answers containing errors fell from 30.5% to 8.0%.

There are weak spots too. On unfamiliar commercial generators the agent is cautious: it labels all real MintVid clips as real, but finds only 64.6% of Sora 2 fakes and 67.0% of Jimeng fakes. The audio tools are not fine-tuned because there is little labeled audio data. The authors warn that the model is built for research and should not make important decisions about people on its own.

The inference code, demo, and LoRA adapters are released under Apache 2.0. The ForenCue-30K dataset and the training code are not published, and AgenticCue is promised for later, so the training cannot be reproduced yet. You can take the model and plug it into a content verification pipeline right now. The most sensible use is as a first-pass filter with manual review, especially for video from new generators, where accuracy is noticeably lower.

Story

Key Questions Answered

How the expert team works?

Imagine an editor who receives a suspicious clip and hands it out to colleagues. One examines skin, hair, fingers, teeth, light, and shadows. A second watches the physics: how bodies move, how liquid flows, what happens when an object passes behind an obstacle. A third judges common sense, such as a cat holding a steering wheel the way a human would. A fourth reads the text. Two more tools listen to the voice and the speech tempo. That is how the system works.

What is new in training?

Earlier agentic detectors, such as ARGUS, trained only the main agent, while the experts stayed frozen. If an expert cannot see an artifact itself, no amount of routing will fix that. Updating both sides at every step is also a bad idea: the main agent decides what the experts will see, and their answers change its decisions, so each becomes a moving target for the other. The authors decoupled the training speeds and called the method DTPO (Dual-Timescale Policy Optimization).

What are the key details on The numbers, the limitations, and what to do about them?

On GenBuster-Bench, the agent scores 97.7% versus 94.3% for VideoVeritas. On GenBuster-Bench++ it scores 97.5% versus 91.4%, and on the six MintVid generators 88.0% versus 82.2%. The binary detectors NSG-VD and ReStraV, trained on the same data, stayed at guessing level: 46.4% and 51.9% on GenBuster-Bench. Gemini-3.1-Pro finds 99% of fakes on GenBuster-Bench++, but it also flags real videos as fake, so its accuracy is 54.3%.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…