Predict how the human brain responds to video content — before you publish it.
NCI uses Meta's TRIBE v2 brain encoding model to predict neural activation patterns from video content entirely computationally. No lab, no subjects, no $150k studies — just a GPU and your video file.
The content industry is reactive: publish, measure, iterate. Traditional neuromarketing is predictive but costs $15,000–$150,000 per study and takes weeks. NCI bridges the gap:
| A/B Testing | Platform Analytics | AI Scoring Tools | Neuromarketing | NCI | |
|---|---|---|---|---|---|
| Timing | Post-publication | Post-publication | Pre-pub (metadata only) | Pre-publication | Pre-publication |
| Evaluates actual video | Indirectly | Indirectly | No | Yes | Yes |
| Cost per evaluation | Ad spend required | Free (post-hoc) | $10–50/mo | $15k–$150k/study | ~$0.10–$1/video |
| Requires audience | Yes | Yes | No | Yes (20–40 subjects) | No |
| Explains why | No | Limited | No | Yes | Yes |
- 1,000+ hours of fMRI training data from ~720 participants
- 20,484 cortical vertices predicted per timestep
- ~2 minutes per 30–60s video on a standard GPU
- 7 interpretable neural engagement signals
- 5 composite engagement metrics
NCI maps raw voxel predictions onto the Yeo 7-network brain parcellation — a standard neuroscience atlas — to produce interpretable cognitive signals:
| Network | Signal | What It Captures |
|---|---|---|
| Visual | Visual Salience | Scene complexity, motion, color contrast |
| Somatomotor | Embodied Response | Sensory/motor resonance, "feeling" the content |
| Dorsal Attention | Sustained Attention | Top-down focus, visual tracking |
| Ventral Attention | Surprise Detection | Novelty, salience breaks, hook moments |
| Limbic | Emotional Resonance | Affective processing, emotional engagement |
| Frontoparietal | Decision Activation | Critical thinking, CTA readiness |
| Default Mode | Narrative Engagement | Story immersion, self-referential processing |
From the 7 network signals, NCI computes five composite metrics:
- Attention Retention Score (ARS) — Does the content sustain focus across its duration?
- Emotional Impact Index (EII) — How strongly does it engage emotional processing?
- Hook Strength Score (HSS) — How effectively do the first 1–3 seconds capture attention?
- CTA Activation Score (CAS) — When are viewers in a decision-making neural state?
- Neural Engagement Score (NES) — Single composite ranking across all dimensions
We analyzed five diverse short-form videos to demonstrate that different content formats engage qualitatively different brain systems — not just different degrees of the same response, but entirely different neural architectures.
 Business Education Somatomotor + Default Mode dominant — voice delivery matters more than visuals |
 Tech/AI News Elevated Frontoparietal — viewers in critical-thinking mode, prime for CTAs |
 UGC Review Highest Somatomotor (27%) — embodied/sensory processing drives authenticity |
 Product Demo Visual + Dorsal Attention dominant — pure "show don't tell" engagement |
 Viral/Satisfying Highest Ventral Attention (17%) — surprise peaks create addictive looping |
Video → TRIBE v2 Brain Encoding → 20,484 Voxel Predictions → Yeo 7-Network Mapping → Engagement Metrics
The temporal dynamics reveal when each brain network activates across the video timeline:
Salience peaks in the Ventral Attention network identify hook moments — the frames that neurally "grab" viewers:
Mean brain activation across the cortical surface:
Each video gets a full neural analysis — network timecourses, activation percentages, hook detection, and engagement scoring:
Business Education (Leila Hormozi)
Tech/AI News (Elon AI Commentary)
UGC Product Review (Perfume Interview)
Product Demo (Sanitary Pad)
Viral/Satisfying (Japanese Ice Cutter)
- Python 3.11+
- Hugging Face account with access to
meta-llama/Llama-3.2-3B(gated model)
chmod +x setup.sh
./setup.shOr manually:
git clone https://github.com/facebookresearch/tribev2.git
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip setuptools wheel
pip install -e ./tribev2
huggingface-cli loginsource .venv/bin/activate
# Text input (easiest smoke test, no video decode needed)
python run_inference.py --text sample_prompt.txt
# Video input
python run_inference.py --video sample.mp4
# Audio input
python run_inference.py --audio sample.wav
# Custom output path
python run_inference.py --video sample.mp4 --output results.npyOutput is a .npy file with shape (n_timesteps, n_vertices) — predicted brain responses on the fsaverage5 cortical mesh, shifted 5s back for hemodynamic lag.
The full research paper is available at paper/neural_content_intelligence.pdf:
Neural Content Intelligence: Using Brain Encoding Models to Predict Social Media Engagement Before Publication
Josh W. — Independent Researcher
NCI is a proof-of-concept. Key limitations include:
- No validation against real-world engagement metrics yet (the critical next step)
- Population-level predictions only (~720 training subjects)
- Training data from controlled lab settings (WEIRD populations)
- Current implementation is visual-only (audio support planned)
See the paper for full discussion.
This project uses Meta's TRIBE v2 model. See the TRIBE v2 repository for model licensing terms.