TRIBE v2 Viral Video Analyzer (amirmushich)

Source: https://x.com/amirmushich/status/2049817739112432097?s=52

Tweet: https://x.com/amirmushich/status/2049817739112432097?s=52

Found via: @telegram on 2026-05-03

Type: Standalone tool / Local web app

Value Score: 3/5

Why not 4: Would need proven correlation between TRIBE-predicted brain activation and actual YouTube Shorts algorithmic performance (currently theoretical), plus confirmation it runs on 8GB VRAM.

Depth: deep

Why deep: The core question—does synthetic fMRI prediction actually improve short-form view counts, or is it neuroscience theater?—requires pressure-testing the mechanism, the hardware feasibility on Eli's RTX 5070, and how it ranks against the existing Claude engagement scorer.

Verdict: Interesting experimental signal, but the leap from "brain activation" to "algorithm performance" is speculative; steal the A/B comparison workflow and timeline visualization concepts rather than adopting the heavy TRIBE v2 inference whole.

What It Is

Amir Mushich built an open-source local web app that wraps Meta's TRIBE v2 (a 1B-parameter trimodal neural encoding model trained on 700+ fMRI subjects) to predict brain responses to short video content. Users upload videos; the app outputs predicted engagement curves, brain-region heatmaps (visual, auditory, language, emotion, memory, prefrontal), editing recommendations, and side-by-side A/B comparison of multiple cuts.

GitHub Stats

The Whole Picture

Standalone tool — a single-purpose analyzer, not a system. It does one job: run TRIBE v2 inference on a video file and surface the predicted neurological response in an editing-friendly UI. The author's tool adds a lightweight recommendation layer (GPT-4o-style suggestions) and multi-video comparison on top of the core TRIBE v2 model. There is no compounding loop or multi-stage workflow beyond "upload → predict → compare."

The Value

As parts

• A/B pre-test comparison workflow → formalize in rav-eade-shorts — The author's side-by-side cut comparison is the most practical feature. Eli's pipeline already generates multiple variants (jump-cut vs. raw, different captions, Ken Burns vs. static). Systematically comparing two exported cuts before upload—using any scoring signal—is a process improvement worth adopting even without TRIBE v2.
• Timeline-based engagement visualization → add to rav-eade-shorts scorer — The idea of plotting a response curve across the video timeline to identify weak moments. Eli's existing scorer (src/engagement_scorer.py, called in pipeline.py Step 4) produces aggregate scores per clip (hook, completeness, flow, shareability, total/100). Adding a simple timeline graph of where attention drops within a clip would make the output more actionable for editors.
• Six-dimension evaluation rubric → expand existing engagement scorer — TRIBE v2 tracks Visual, Auditory, Language, Emotion, Memory, and Prefrontal activation. These six dimensions could inspire a richer Claude-based scoring rubric in the existing pipeline, giving more granular feedback than the current four categories.
• Neurological scoring as second opinion → rav-eade-shorts research branch — If the model can be made to run on Eli's hardware, the predicted brain-activation signal is orthogonal to the existing viral-heuristics approach. It would not replace Claude's engagement scorer but could act as a novelty filter for clips that "should" work neurologically even if they don't fit the standard viral categories.

Relevant Projects

• Rav Eade Shorts — The primary and only relevant project per user guidance. The A/B comparison concept and timeline visualization directly improve the short-generation workflow. The neurological prediction layer is an experimental add-on to the existing Claude-based engagement scorer and jump-cut pipeline.
• Thumbnail Generator — Marginal relevance: TRIBE v2 processes visual frames, so in theory it could score thumbnail images for visual-cortex activation, but the tool is video-native and this is speculative.

Already Doing This?

Eli's pipeline already has a sophisticated engagement evaluation layer. pipeline.py Step 4 calls src.engagement_scorer.score_engagement(), which scores each clip on hook, completeness, flow, and shareability with a 0–100 total and a verdict ("Strong viral candidate", "Good but not great", etc.). This is qualitatively different from TRIBE v2: the existing scorer uses Claude to apply a viral-DNA rubric derived from top-performing shorts, while TRIBE v2 predicts synthetic fMRI activation based on neuroscience training data. They are complementary signals, not replacements.

What Eli does not have:

• Pre-upload A/B comparison of multiple cuts
• Timeline-level engagement curves within a clip
• A neuroscience-grounded evaluation dimension

Community Sentiment

Too new for organic signal. The repo dropped as an MVP in early May 2026. No significant Hacker News or Reddit discussion found yet. The publicimpact.ai team has a similar commercial wrapper and reports strong correlation between predicted brain activation and ad recall, but this is for brand advertising, not YouTube Shorts algorithmic distribution. The buzz around TRIBE v2 itself is mostly research-community enthusiasm (Algonauts 2025 winner) repurposed by marketing blogs. Buzz Assessment: Too new.

Technical Assessment

• Model specs: TRIBE v2 is a ~1B parameter trimodal transformer. It processes video frames, audio, and text in parallel to predict BOLD signals across ~20,000 cortical vertices.
• Hardware reality: The publicimpact.ai implementation runs inference on an A100 GPU. Eli's machine has an RTX 5070 Laptop with 8GB VRAM. For a 1B-parameter multimodal video model, 8GB is likely insufficient or borderline at best—especially processing video frames sequentially. A Rust inference port exists (tribev2-rs) suggesting the community is already trying to optimize for lower-end hardware, but this is early.
• Windows compatibility: Likely Python-based (Flask/Streamlit-style local web app). No Mac-only dependencies expected, but the official facebookresearch/tribev2 repo may assume Linux for some utilities.
• Stability: MVP v1. Expect rough edges, missing docs, and setup friction.

How to Apply This

Parts to steal

1. Build an A/B comparison harness in rav-eade-shorts — Before investing in TRIBE v2 inference, add a --compare flag to pipeline.py that takes two output directories (e.g., jumpcut/ vs. raw/) and runs the existing engagement scorer on both, outputting a side-by-side markdown report. This captures 80% of the author's workflow using tools you already have.
2. Add timeline visualization to engagement scores — Modify src/engagement_scorer.py (or the scorer output formatter) to emit a simple ASCII or matplotlib graph of predicted engagement across the clip timeline, using Claude to score 5-second windows. This surfaces "weak moments" the same way TRIBE v2's ROI curves do, without the heavy model.
3. Test TRIBE v2 inference on a single short — If Eli still wants the neuroscience signal, try running the official facebookresearch/tribev2 model (not the MVP wrapper) on one 30-second Rav Eade clip to verify it loads on the RTX 5070 8GB. If it OOMs, the question is answered and the idea is blocked on hardware.
4. Expand the engagement rubric — Add Memory and Emotional Resonance as explicit scoring dimensions in the existing Claude engagement prompt, inspired by the TRIBE v2 ROI taxonomy. This costs nothing and immediately deepens the current scorer.

Research leads

• TRIBE v2 + thumbnail scoring — If the model runs locally, a static-frame inference mode could score thumbnails for visual-cortex activation before compositing. Not useful until the video pipeline is proven.
• PublicImpact.ai API as a lightweight alternative — If local inference fails on 8GB VRAM, their hosted API might offer per-video TRIBE analysis without hardware investment. Worth checking pricing if the A/B comparison concept proves valuable.

Bottom Line

Do not adopt the full TRIBE v2 inference stack yet because the hardware requirement is uncertain on 8GB VRAM and the brain-to-algorithm correlation is unproven. The single best part to extract is the A/B comparison + timeline visualization workflow—which can be built immediately using the existing Claude engagement scorer.