Gaze Follow Video: a Website Character Whose Eyes Follow the Pointer
Build a footer, hero or 404 page where a character watches the visitor: SandBase generates a still character whose pupils roll in one circle, every frame is calibrated to a gaze angle, and the page seeks the paused video to the frame that looks at the pointer, then looks back when the visitor stops.
Bring a character idea (or your own image or clip), the layout and your brand, copy, links and language. The Skill returns a static page (index.html), a section to paste into an existing site, a React component, the calibrated clip and colours taken from the character.
How it works
With your go-ahead (about $0.80 per attempt), SandBase generates the character image and a 6-second clip in which only the pupils roll once around the eyes.
Calibration finds the eyes, tracks the pupils in every frame and writes an angle-to-frame table; clips that blink, stop short of a full circle or move the body are rejected with the reason.
The page keeps the clip paused and seeks to the frame looking at the pointer; every frame is a keyframe, so each seek decodes one frame and the eyes react at once.
Frequently asked questions
Does the page call a model?
No. The finished page is static files: HTML, one script, fonts and two clips. Models are only used when the Skill generates a new character.
What does a new character cost?
About $0.80 per attempt through SandBase (image $0.075, 6-second clip $0.72). Roughly half of the clips are rejected by calibration, so budget for a second attempt.
Can I use my own character or clip?
Yes. Pass your own image to skip the image step, or calibrate your own clip. It must hold the body and camera still, keep the eyes open and show dark pupils on light eyes.
Why do the links do nothing here?
The example runs in a sandboxed frame with its links disabled. In your own page they go wherever you set them in site.json.
Does it work on phones?
Yes. Narrow screens loop a small clip; a tap loads the full clip and makes the character look at the tap. Checked in a 390 px phone viewport.
Why is there no online generation?
Building a new scene needs a model key inside the sandbox, which the backend doesn’t support yet. Run the Skill with your own agent for now.