gazeportrait

a sprite atlas portrait that follows your cursor. one image, one canvas, zero dependencies.

try it

git clone https://github.com/agenticnotetaking/gazeportrait.git
cd gazeportrait
open example/index.html

an example atlas is included — it works immediately. the demo shows:

• the portrait tracking your cursor horizontally
• a frame bar mapping cursor position to the active frame
• a grid minimap highlighting the current atlas cell
• a hatched deadzone marking the upper hemisphere (no vertical tracking)

walkthrough

1. generate or record a head-turn video

you need a short video (3–10 seconds) of a head turning left to right. options:

• AI generation — use Kling, Runway, Pika, or Stable Video Diffusion. prompt a smooth left-to-right head turn from a single portrait image.
• real footage — film someone (or a pet) turning their head. works the same way.

2. extract frames

./scripts/extract-frames.sh your-video.mp4

this extracts frames at 30 fps as PNGs into frames/ using ffmpeg.

3. find the clean sweep

pip install -r scripts/requirements.txt
python scripts/build-contact-sheet.py frames/ --cols 15

open contact-sheet.jpg. you'll see a numbered grid of every frame. find the range where the head sweeps cleanly in one direction — note the start and end frame numbers.

4. build the atlas

python scripts/build-atlas.py frames/ \
  --start 120 --end 270 --count 49 --cols 7 \
  --size 256x256 --out example/atlas.jpg

• --start / --end — the frame range of your clean sweep
• --count 49 — samples 49 frames evenly across the range (7x7 grid)
• --reverse — add this if the sweep goes right-to-left in the video

5. open the demo

open example/index.html

edit the config block at the top of index.html to match your atlas dimensions if they differ from the defaults.

advanced pipeline (human faces only)

one command — uses MediaPipe face detection + OpenCV head-pose estimation to automatically sort frames by yaw, pick the sharpest per grid cell, and fill gaps.

pip install -r scripts/requirements-advanced.txt
python scripts/build-sprite-sheet.py your-video.mp4 --out example/atlas.jpg --grid 7

this doesn't work for pets or non-human subjects — use the manual pipeline above instead.

how it works

a canvas draws one frame at a time from a sprite atlas (a single jpeg containing all frames in a grid). cursor X position maps to a frame index. smoothing prevents frame-snapping:

smooth += (mouse - smooth) * 0.35
drift = sin(elapsed / 3000) * 0.015

the 0.35 controls responsiveness — higher values track the cursor more tightly. the drift adds a subtle idle sway so the head never fully stops.

config

both example/index.html and react/GazePortrait.tsx have a config block at the top:

| param | default | what it does | |-------|---------|--------------| | COLS | 7 | columns in the atlas grid | | ROWS | 7 | rows in the atlas grid | | FRAME_W | 256 | frame width in px | | FRAME_H | 256 | frame height in px | | INVERT | false | flip direction (if atlas sweeps R→L) | | ATLAS_SRC | atlas.jpg | path to your sprite atlas |

tuning the feel

| value | effect | |-------|--------| | 0.35 → 0.5 | snappier, more direct tracking | | 0.35 → 0.2 | more inertia, floatier feel | | drift * 0.015 → 0.03 | more idle sway | | drift * 0.015 → 0 | completely still when idle |

limitations

• horizontal only — cursor Y position has no effect. the atlas is traversed as a 1D sequence. the demo shows a hatched deadzone over the upper hemisphere to indicate this.
• manual sweep selection — for non-human subjects, you identify the clean sweep visually from the contact sheet. the advanced pipeline automates this for human faces.
• frame order matters — the simple pipeline packs frames in video order. if the sweep goes right-to-left, use --reverse.

toward 2D tracking

the advanced pipeline already builds a 2D yaw x pitch grid — columns are yaw, rows are pitch. the viewer doesn't use the second axis yet.

full 2D tracking would map cursor X → column (yaw) and cursor Y → row (pitch):

const col = Math.round(mouseX * (COLS - 1));
const row = Math.round(mouseY * (ROWS - 1));

this needs a source video with full head orbit (not just a horizontal pan) and a larger grid (12x12+) for smooth coverage.

what's in here

example/
  atlas.jpg                    example atlas (works out of the box)
  index.html                   vanilla JS demo with debug overlay

react/
  GazePortrait.tsx             drop-in React component

scripts/
  extract-frames.sh            ffmpeg → PNGs
  build-contact-sheet.py       numbered review grid
  build-atlas.py               frame range → sprite atlas (with even sampling)
  build-sprite-sheet.py        video → atlas in one step (advanced, human only)
  requirements.txt             Pillow
  requirements-advanced.txt    opencv, mediapipe, numpy

license

MIT