The guide · setup guide
The viral engine.
Engineer a viral video using AI that simulates how your brain reacts to every second of it. Meta's research team just open-sourced TRIBE v2 — a model trained on 1,115 hours of fMRI brain scans that predicts, second by second, when a viewer is paying attention and when they're about to scroll. This is the full step-by-step setup, plus how to connect it to Claude to compare up to 10 versions and find the winner. Zero coding required.
The full setup
🔒 Drop your email to reveal the full setup
Get the full setup, free.
Drop your email and every step, cell, and prompt unlocks instantly, plus every other guide and resource in the library. One email a week, unsubscribe anytime.
What is TRIBE v2?
Meta's AI research team (FAIR) just open-sourced a model called TRIBE v2. It was trained on 1,115 hours of fMRI brain scans from 720 real volunteers — people who watched movies, listened to podcasts, and read text while their brain activity was recorded at high resolution inside an MRI machine.
TRIBE v2 learned the patterns. Now it can predict how a human brain responds to any video — without putting anyone in an MRI machine. You give it a video; it tells you, second by second, which brain regions are lighting up and which are going dark:
- When attention is spiking
- When engagement flatlines
- When a moment gets encoded into memory
- When the viewer is about to scroll
Meta's own research shows TRIBE v2's predictions are more accurate than a single real brain scan — because real scans are noisy (heartbeats, movement, device artifacts) and the model strips all that out. They open-sourced everything, and it's completely free.
Paper: TRIBE: TRImodal Brain Encoder — arXiv 2507.22229. Meta blog: Introducing TRIBE v2. License: CC BY-NC 4.0 (free for non-commercial use).
What you'll need (all free)
| Tool | What it's for | Cost |
|---|---|---|
| Google Colab | Runs the model in the cloud — no computer required | Free (Pro recommended) |
| Hugging Face account | Download the model weights | Free |
| Claude | Compare your video versions and tell you the winner | Free tier works |
| Your video file | What you're analysing | — |
Part 1: One-time setup (15–20 min)
If you're a visual learner you can also follow this simple video guide.
Open the official Colab notebook
Open Meta's official demo notebook. This is Meta's own notebook — you don't need to write any code, every cell is already written for you.Switch to a GPU
The model needs a GPU. In the top menu click Runtime → Change runtime type, and under "Hardware accelerator" select T4 GPU, then click Save. If you later see out-of-memory errors, upgrade to Colab Pro ($10/month) and select an A100 GPU instead — the full TRIBE v2 model is large (≈30GB of weights loaded at once). T4 works for shorter videos; A100 is more reliable.Create a Hugging Face account
Go to huggingface.co and sign up for a free account. After signing in, click your profile picture → Settings → Access Tokens, click New Token, give it any name (e.g.tribe-token), set permission to Read, click Create token and copy it — you'll need it in a minute.Request Llama model access
TRIBE v2 uses Meta's Llama 3.2 model to process any text in your video, and Llama is gated. Go to the Llama 3.2 model page, fill out the terms form and submit, then wait for approval — usually 30 min to 2 hours. You'll get an email when you're approved. You only need to do this once; after approval your token automatically has access.Add your token to Colab
Back in your notebook, click the 🔑 key icon (Secrets) in the left sidebar, click Add new secret, typeHF_TOKEN in the "Name" field, paste your Hugging Face token in the "Value" field, and toggle on Notebook access.Install the model
Find the first code cell (below), click its ▶ play button to install all required packages. When it finishes you'll be asked to restart the runtime — click Runtime → Restart session, and do NOT run the install cell again after restarting.The install cell:
Part 2: Running your first video
Step 7 — Load the model
Run the next cell. This downloads the model weights and gets everything ready — it takes 2–5 minutes the first time.
Step 8 — Upload your video
The notebook includes a sample video by default. To use your own video, find the cell that downloads the sample video and replace it with this:
Run that cell — an Upload button appears. Click it and select your video file.
Supported formats: MP4, MOV, AVI. Recommended: keep videos under 3 minutes for faster processing on free Colab.
Step 9 — Run the brain analysis
Run these cells in order. This is the core step — TRIBE v2 is now predicting how a brain responds to every second of your video.
Note: TRIBE v2 outputs predictions offset by 5 seconds — this accounts for the natural delay between a stimulus and the brain's blood-flow response (hemodynamic lag). So second 10 in the output = the brain's reaction to second 5 of your video.
Step 10 — Visualise the results
This generates an interactive 3D brain heatmap synced to your video:
- Red/warm colours = high activation (brain is engaged)
- Blue/cool colours = low activation (brain is disengaged)
- The map updates second by second as the video plays
Key brain regions to watch:
| Region | Location | What it means for your video |
|---|---|---|
| Dorsal attention network | Top/parietal | Viewer is actively paying attention |
| Visual cortex | Back of brain | Processing what they're seeing |
| Auditory cortex | Sides/temporal | Responding to sound/music |
| Default mode network | Spreads across midline | When this activates, mind is wandering — you're losing them |
| Hippocampus area | Medial temporal | Memory encoding — they'll remember this moment |
Step 11 — Get second-by-second numbers
To get numerical data you can feed into Claude, run this and copy the output:
Part 3: Connect to Claude — find your best version
This is where it gets powerful. You can run up to 10 versions of your video through TRIBE v2, collect the second-by-second activation data from each one, and then ask Claude which one wins and exactly why.
For each video version, run Steps 8–11 and copy the output table. You'll end up with something like this:
Then paste this prompt into Claude:
Claude comes back with a ranked analysis, tells you the winner, identifies the exact seconds where each version loses attention, and gives you concrete re-edit instructions.
Part 4: Re-edit based on what the brain says
Once Claude identifies your drop-off points, here's how to fix them:
| Problem the brain data shows | What to do in your edit |
|---|---|
| Activation drops in first 3s | Re-cut your hook — lead with the most visually striking frame you have |
| Sustained low activation in middle | Add a pattern interrupt: cut to a new angle, add text overlay, use a sound effect |
| Activation never fully recovers after drop | Remove the dead section entirely — it's killing the rest |
| Activation high but drops at the end | Your CTA is weak — add urgency, movement, or a direct question |
| Default mode network activating | Your visuals are too static — add motion, zooms, or jump cuts |
| Audio regions flat | Your audio is too monotone — vary pace, add music, or cut silence |
Run the re-edited version back through TRIBE v2, compare the new activation data against the original, and repeat until the brain stays engaged.
Quick reference links
Colab notebookOpen in Colab ↗GitHub repofacebookresearch/tribev2 ↗Modelfacebook/tribev2 on Hugging Face ↗Llama 3.2 accessmeta-llama/Llama-3.2-3B ↗Interactive demoaidemos.atmeta.com/tribev2 ↗Research blogai.meta.com/blog/tribe-v2 ↗Paperarxiv.org/abs/2507.22229 ↗
Common issues & fixes
- "CUDA out of memory" error — your GPU ran out of VRAM. Upgrade to Colab Pro and switch to A100 GPU. The full model needs ~30GB.
- "Access to model is restricted" on Hugging Face — you haven't been approved for Llama yet. Wait for the approval email (up to 2 hours) or check your Hugging Face inbox.
- "HF_TOKEN not found" error — go back to the 🔑 key icon in the Colab sidebar and make sure you added the secret with the exact name
HF_TOKEN(all caps, underscore). - Upload button doesn't appear — make sure you commented out the sample video download lines before running the upload cell.
- Predictions look the same for every second — your video might be too short (under 10 seconds) or the model needs more frames. Try a video that's at least 30 seconds.
- Runtime disconnects mid-run — Colab free tier disconnects after 90 minutes of inactivity. Upgrade to Pro or keep the tab active. The model weights are cached in
./cache, so you don't need to re-download them if you reconnect.
What the numbers actually mean
TRIBE v2 outputs values on the cortical surface — each number represents a predicted BOLD signal (Blood Oxygen Level Dependent), which is what an fMRI scanner measures.
- Higher values = more oxygenated blood flowing to that brain region = more neural activity = more engagement
- Lower values = less activity = the brain is in a resting or wandering state
You don't need to understand neuroscience to use this. The pattern is simple: watch for the dips. Every time the mean activation drops below the baseline for your video, that's a second your viewer's brain is checking out. The goal is a graph that stays high, climbs at key moments, and never flatlines.
Original guide by Artem Novitckii. More from him on Instagram, in the Skool community, and across his resource library.