Rendering · Nov 2026
Listening is visible: why micro-expression matters more than resolution
People judge whether they're being heard from small, fast facial signals. We designed Lumen around them.
When people describe an AI face as uncanny, they rarely mean it looks low-resolution. They mean it doesn't seem to be listening. Between turns, a real listener nods slightly, narrows their eyes, glances away while thinking and comes back.
Lumen treats those in-between moments as first-class output. While the user is speaking, the renderer receives a stream of listening cues from Cadence (is this a thought pause or a finished turn?) and from Iris (is the speaker animated, hesitant, upset?) and blends a small library of attentive behaviours accordingly.
In internal blind comparisons, participants consistently preferred lower-resolution renders with active listening over higher-resolution renders that held still. We now allocate rendering budget to motion fidelity around the eyes and mouth before pixel count.
This note describes our design approach; a full evaluation write-up will accompany the Lumen-1 public release.