How the Ninox Model Works

Why AI Video Detection Matters

Artificial intelligence can now create video of people saying things they never said, doing things they never did, or being in places they never were — and it takes only seconds to produce. These are called AI-generated videos, and they are becoming indistinguishable from the real thing.

The consequences reach far beyond social media. A fabricated video of a politician, a CEO, or a public figure can move financial markets, trigger political crises, destroy reputations, and undermine courtroom evidence. In an age where "seeing is believing," that old saying has become a liability.

Detection tools are the countermeasure — the forensic equivalent of carbon-dating a document or fingerprinting a suspect. But a detection tool that cannot explain its reasoning is only half the solution. Courts, journalists, and platform moderators need more than a verdict. They need evidence they can examine.

The Base Model: Ninox 1

85.12% accuracy · three core components

Ninox 1 is built from three components that work in sequence: a Latent Encoder, a Patch Encoder, and a Visual Transformer. Together they take a raw video and produce a verdict. Understanding each one also reveals why the model is well-suited to this task — it is deliberately designed to understand the same internal language that AI video generators use to create synthetic footage.

01

Latent Encoder

A raw video file contains an enormous amount of data — millions of pixel values per second. Working with all of that directly would be like trying to analyse a book by examining every individual ink droplet on every page. The Latent Encoder compresses each frame into a much smaller, information-dense representation called a latent.

Think of a latent as the mathematical essence of a frame — not the pixels themselves, but a compact description of what the frame contains: shapes, textures, relationships, structure. This compression discards the redundant parts and retains the meaningful ones. This approach follows the same principle that underpins AI video generation: rather than working with raw pixels, operate on compact representations that emphasise structure over surface detail. Ninox 1 adopts this philosophy — not because it has access to the internal workings of any specific generator, but because this level of representation is where the structural traces of synthesis tend to be most legible.

02

Patch Encoder — and Tubelets

Once the video exists in latent space, the Patch Encoder divides each frame into a grid of small, equally-sized tiles called patches. The relationship between a patch and an image is the same as the relationship between a word and a sentence: just as a sentence is a sequence of words, an image is a grid of patches. This makes them easier for a transformer-based model to reason about — the same fundamental approach that powers large language models, applied to video.

But video adds a dimension that a static image does not have: time. To capture it, the Patch Encoder groups patches into units called tubelets. A tubelet takes the same patch — the same x/y position in the frame — from two consecutive frames and bundles them together. This means every unit the model analyses contains not just what a region looks like, but how it has changed from one moment to the next.

This is significant for detection. A generated video may have patches that look convincing in isolation. But the tubelet — pairing each patch with its twin from the next frame — gives the model access to how that region changed over time. Patterns that appear consistent frame-by-frame can reveal subtle irregularities when their temporal evolution is encoded directly. By capturing change alongside appearance, the Patch Encoder exposes a dimension of evidence that frame-by-frame analysis would miss entirely.

The tubelet concept follows the approach described in the ViViT (Video Vision Transformer) research paper, which established this as an effective way to encode both spatial and temporal information simultaneously.

03

Visual Transformer

After the Patch Encoder produces a sequence of tubelets, the Visual Transformer has one job: understand how they all relate to each other. It does this through a mechanism called attention — at each step, it asks "which other parts of this video are most relevant to understanding this part?"

A skilled human reviewer watching for a deepfake does something similar: they do not just stare at one spot. They notice that the way a person's hair moves does not match the wind visible in the background; that the shadow on a face points in a different direction from the shadow on a wall three metres away; that a mouth shapes a word that the rest of the face does not. The Visual Transformer does this systematically, across every tubelet in the video simultaneously.

This cross-video awareness is where AI-generated footage most often betrays itself. A generator constructs a video region-by-region, and the local results can each look plausible. But global consistency — the kind that emerges naturally from a real camera recording a real scene — is genuinely difficult to synthesise. The Visual Transformer is specifically designed to surface those global inconsistencies.

The Gap: What Ninox 1 Still Misses

Even at 85% accuracy, roughly 1 in 7 videos is misclassified. The core limitation comes down to the scope of what tubelets capture. A tubelet pairs the same patch across two consecutive frames — giving the model a window onto short-range change. What it does not directly capture is continuous motion across many frames: the arc of a hand gesture, the sustained trajectory of a moving object, the cumulative drift of a camera pan.

In real footage, movement is governed by physics — weight, momentum, inertia. Objects accelerate and decelerate in predictable ways; bodies carry the momentum of their last position into the next. In AI-generated footage, this continuous physical logic is often subtly wrong. The movement may look plausible frame-to-frame, but over a longer window it drifts in ways that no real object would. Ninox 1's two-frame tubelet window is not wide enough to consistently catch this kind of sustained irregularity.

This is the gap that the next version of the model is specifically designed to close.

The Improvement: Adding Optical Flow

Ninox 1.1-Flow — 93.53% accuracy

Optical flow is a technique for computing, at every point in a frame, exactly how much and in which direction each pixel has moved since the previous frame. The result is not another image — it is a dense map of motion vectors covering the entire scene, updated at every moment in the clip. Where a normal frame answers "what does this look like?", an optical flow map answers "how is everything moving, and at what speed?"

The maps are generated by a dedicated flow estimation algorithm applied to each consecutive frame pair as a preprocessing step — before the model processes anything. For each pair, the algorithm produces a dense field: two values per pixel, encoding horizontal and vertical displacement. The result is a separate motion input for every moment in the clip.

Imagine watching a river. A photograph tells you the water is there. But a current map — showing exactly how fast the water moves, in which direction, at every point — tells you something fundamentally different. Optical flow gives Ninox 1.1-Flow that current map for every moment of the video.

Ninox 1.1-Flow adds a second parallel stream to the original architecture. The first stream is unchanged: raw frames through the Latent Encoder, Patch Encoder, and Visual Transformer, producing a classification score from appearance and structure. The second stream takes the pre-computed flow maps and runs them through an identical pipeline — its own Latent Encoder, Patch Encoder, and Visual Transformer — trained independently on motion evidence alone. A learned fusion layer then combines the two scores into a single verdict, having learned during training how to weight each stream depending on the type of evidence present.

The addition of optical flow lifted accuracy from 85.12% to 93.53% — an improvement of more than 8 percentage points. In practical terms, the upgraded model correctly handles roughly 57 more cases out of every 100 that the base model would have misclassified.

Why Evidence Matters, Not Just a Verdict

Many AI detection tools operate as a black box: you put a video in, you get "real" or "fake" out, and you are expected to trust the answer. That is not good enough for any situation where the verdict carries real-world consequences.

The Ninox system can generate an attribution video alongside its verdict. An attribution video is a copy of the original footage with each region highlighted according to how strongly it influenced the model's decision. Regions that drove the prediction glow brightly; regions that had little bearing on the outcome remain dim.

Evidence unlocks four things that a bare verdict cannot:

The Cost of Getting It Wrong

Even at 93.53% accuracy, roughly 1 in 15 videos will receive the wrong verdict. Understanding the type of error matters as much as the error rate itself. There are two distinct failure modes, and they carry very different consequences.

Neither failure mode is acceptable in high-stakes situations. This is why accuracy improvements — from 85% to 93% — are not incremental refinements. Each percentage point represents real decisions, real content, and real people.

It is also why the attribution evidence layer is not optional. When a misclassification occurs, the evidence trail is what allows it to be challenged, corrected, and learned from — rather than silently accepted.