Platforms are declaring war on AI slop. Snapchat, YouTube, LinkedIn, Substack — all rolling out detection, labeling, suppression. Good. Necessary. But as developers, we should ask: what happens to the human runtime when the input stream gets cleaner but denser? Visual attention isn't infinite. Binocular coordination, saccadic accuracy, vergence stamina — these are hardware constraints. And for ~3 % of users (amblyopia) plus a much larger cohort with sub‑clinical visual fatigue, the bottleneck is real.
Amblyotube is a Meta Quest app we built at Seven Sportz. The core mechanism is stereoscopic split rendering: a different YouTube video stream is sent to each eye, forcing cortical fusion during passive viewing. The stack runs on Unity with the Meta XR SDK, and the clinical principle — dichoptic training — is delivered through consumer content that people already watch.
Dominant‑Eye Shader – a fragment shader that modulates the strong eye’s texture. It exposes a configurable blur kernel, contrast/brightness curves, and an alpha opacity parameter. At alpha = 1 the shader acts as a digital patch; at partial values it becomes a binocular‑balance tool that can be tuned per session.
Lazy‑Eye Sharpener (MFBF) – MediaPipe/YOLO‑style person detection runs on the CPU/GPU, produces a mask, and a sharpening convolution is applied only to the lazy eye’s render target. The pipeline sustains 72–90 fps depending on headset model, keeping latency low enough for comfortable prolonged use.
Magenta Focus Cue – a UI overlay rendered exclusively to the lazy‑eye view. The colour is a non‑spectral magenta chosen to avoid collisions with typical video backgrounds, providing a persistent attentional anchor without visual clash.
Visual Accents – post‑process outline (Sobel on the person mask) and highlight (bloom on the mask) are driven by a sine‑wave pulse controller at a configurable frequency, giving a subtle rhythmic cue that encourages sustained fixation.
Session Guardrails – recommended 30–40 min per session, hard cap at 60 min. Lazy‑eye selection is stored in persistent settings; an incorrect selection inverts the therapeutic vector, so the onboarding flow forces a verification step.
Amblyotube is not a game and not a “wellness” toy. It is classified as training/assistive software, not a medical device, and is positioned to complement professional vision‑care programs. The app demonstrates that the receiver side of the content pipe — human visual processing — is the next frontier for XR, accessibility, and human‑computer interaction research.
https://www.meta.com/en-gb/experiences/amblyotube/25906906972338493/
Top comments (0)