Eye Contact Correction: Redirecting the eyes to look at the camera

Original Article ↗ Hacker News Discussion ↗

Perceived quality & limitations

  • Many find the demo technically impressive and fast, with better results than older gaze-correction tools.
  • Others note the sample is mild (eyes already near camera); they want examples with large head turns and “normal” movement and for the system to stop correcting in extreme poses.
  • Some say other vendors (Apple, Google, Nvidia) are more conservative, correcting only within a limited gaze range, which feels more natural.

Comfort, naturalness & uncanny valley

  • Several people find the corrected video more uncomfortable or “creepy” than the original, especially due to:
    • Overly fixed stare and lack of saccades.
    • Continuous eye contact that feels like an interrogation or horror-movie portrait.
  • Suggestions: enable randomized “look away” behavior by default, track blinks and micro-movements, and avoid 100% constant eye contact.

Ethics, honesty & social signaling

  • Strong split:
    • Some see correction as “lying” about attention and presence, undermining cues managers/teachers/spouses use to judge engagement.
    • Others argue the uncorrected view is the lie, since people are genuinely looking at the screen/other person but appear to be looking away because of camera placement.
  • Concerns that masking disengagement will worsen remote-work trust, hiring fraud, and leadership feedback loops.
  • Some neurodivergent people worry about pressure to use such tools to hide traits like avoiding eye contact.

Use cases, demand & pricing

  • Main use case cited: videoconferencing, interviews, and remote work where eye contact is valued.
  • Some users say they never missed this feature and prefer natural gaze.
  • Pricing (e.g., $0.10/minute) is criticized as too expensive; local GPU-based tools (e.g., Nvidia Broadcast/SDK) are preferred for everyday calls.

Alternatives & future directions

  • Hardware approaches: teleprompter-style mirrors, drop-down/arm cameras, cameras behind/inside displays, beam-splitters.
  • Ideas for more advanced systems:
    • Virtual cameras that re-render the whole face from a new viewpoint.
    • Gaze correction relative to the on-screen position of the person you’re looking at.
  • Broader worries about normalized AI video manipulation, deepfakes, evidence authenticity, and possible future gaze-tracking/advertising abuse.