A virality score before you post
Higgsfield's Virality Predictor scores a clip from a modelled brain response. What the score can tell you, what it cannot, and why real audiences still decide.
Every social team would like to know, before posting, whether a clip will travel. Higgsfield, the AI video platform, now offers an answer among its apps. Upload a clip of up to 15 seconds and its Virality Predictor returns a virality score, the second where the hook peaks, a hold rate and a heatmap of the brain regions the clip is said to activate.
The idea is appealing and the screen looks scientific. Here is what Higgsfield claims, what it has published about how the tool works, what research says about predicting virality, and how a score compares with testing on real people.
What it claims
The product page promises “a modeled read on whether your clip earns the scroll or loses it”. The method is described in one sentence: “A modeled audience watches while we map the brain response across vision, sound, memory, attention, language.” The heatmap, in the page's words, shows “exactly which regions of the brain your clip is firing”.
The score is pitched with confidence: “Not a guess, not a vibe check”, and “The closest thing to a test screening before you hit publish.”
What is published, and what is not
Higgsfield does not say which model sits behind the score, what data it was trained on, or how its scores compare with the real views, watch time or shares of published clips. We found no accuracy figures and no validation study. Until those exist, the claim that a score predicts virality is unproven.
The vocabulary points to a known field of research. Brain encoding models learn to predict how the brain responds to video from scans of people watching videos. TRIBE, published in July 2025, predicts whole brain fMRI responses to video and won the Algonauts 2025 competition. Such models estimate how an average brain region reacts. That is a different question from whether strangers will stop, watch and share.
There is evidence that real brains carry useful signals. In a study published in PNAS in 2020, researchers scanned 40 people as they watched videos. Activity in two regions at the start of each video forecast how often and how long the same videos were watched on YouTube, beyond conventional measures. That result came from real scanners and real people, on a few dozen videos. A modelled audience is one step further removed.
Why virality resists prediction
The deeper problem is that success spreads socially. In a well known experiment published in Science in 2006, 14'341 participants downloaded unknown songs, some of them able to see what others had chosen. Social influence made success both more unequal and less predictable. The authors wrote that “The best songs rarely did poorly, and the worst rarely did well, but any other result was possible.”
A study of Twitter cascades presented in 2016 found that even the best models, with extensive data on users, content and past performance, explained “less than half of the variance in cascade sizes”, and argued that even unlimited data would not remove that limit. What does work is watching early spread. An earlier study found that whether a cascade keeps growing becomes more predictable “as we observe more of its reshares”. The best predictor of a clip's reach is its first real audience.
What a score can and cannot do
Used carefully, a score like this can still help. It can flag a slow opening, show the second where attention may drop, and serve as a quick first filter between two cuts of the same clip. As a prompt for a closer look, it costs little.
It cannot know your audience, your timing, what else is in the feed that day, or whether the hook fits your brand. A virality score is not a probability and not a forecast of views. And if every team optimises for the same score, every feed fills with the same hooks, which in our view is the opposite of what makes a clip stand out. Even Higgsfield writes that more output “doesn't decide the outcome on its own”.
Testing on real people
Real tests are now built into the platforms. Meta's A/B tests show each version of an ad to a separate part of the audience and make sure “nobody sees both”, then compare results on cost per result. YouTube lets creators test up to three titles and thumbnails on one video, judges them by watch time share and names a winner only when the difference is statistically significant. Both measure what actually happened.
That is how Tempo is set up. Every idea is created, tested and checked for its chances before a person approves it, and we treat that check as a filter, not a forecast. What decides the next month is how the real posts performed, which flows back into the plan. A virality score is a good question to ask of a clip. The audience gives the answer.
