Why Software Video Stabilization Fails on Long-Term Timelapses (And How We Solved It in the Viewfinder)

If you have ever tried creating a 12-week gym progress video, a year-long daily selfie time-lapse, or a seedling-to-bloom reel, you know the frustration: you take 90 photos, drop them into a video editor with digital stabilization, and the resulting video looks warped and jittery.

Here is why post-capture software stabilization fails for long-term progress photos, and how an optical ghost overlay approach solves it.

---

1. The Physics of Perspective & Parallax

Digital post-capture stabilization works by detecting high-contrast feature points across consecutive frames, calculating an affine or homography transformation matrix, and translating/rotating the image to align those points.

This works brilliantly for 60fps continuous video because camera movement between frames is tiny (fractions of a millimeter).

However, in multi-month progress photography:

• Lens Distance Drift: Standing 1.8m away in Week 1 vs 2.1m away in Week 4 changes the apparent size of the subject.
• Perspective Parallax: If your camera is tilted 3° upward on Day 10 and 2° downward on Day 20, the relative distance between background objects and foreground facial landmarks shifts geometrically.
• The Crop Penalty: Post-stabilization must crop the overlapping convex hull of all frames, losing up to 30% of the image resolution.

No 2D affine transform can undo 3D optical parallax after the photo has already been taken.

---

2. Solving Framing Before the Shutter Fires

To eliminate perspective distortion, the alignment must happen inside the live viewfinder before the sensor exposes the frame.

We built SameShot, an offline-first camera application and browser utility suite that projects your previous photo as a semi-transparent ghost overlay directly in the live camera feed:

1. Draggable Opacity Slider: Users adjust the ghost layer (0–100%) to match their subject against the prior frame.
2. On-Device ML Auto-Capture: Using local Google ML Kit pose and face landmark detection, the shutter fires hands-free the millisecond the subject hits alignment thresholds.
3. Zero Network Calls: The entire pipeline runs locally in browser memory and on-device without sending any image data to a cloud backend.

---

3. Building a Client-Side Difference Map in Vanilla JS

To help users inspect whether two photos match in geometry and lighting, we built a zero-server Photo Alignment Checker using the HTML5 Canvas 2D API:

javascript
// Local Difference Blend Mode via Canvas
const canvas = document.createElement('canvas');
const ctx = canvas.getContext('2d');

canvas.width = baseImage.width;
canvas.height = baseImage.height;

// Draw baseline photo
ctx.drawImage(baseImage, 0, 0);

// Set difference composite operation and draw new photo
ctx.globalCompositeOperation = 'difference';
ctx.drawImage(comparisonImage, 0, 0);

// Invert to highlight drift: identical pixels become pure white
ctx.globalCompositeOperation = 'difference';
ctx.fillStyle = 'white';
ctx.fillRect(0, 0, canvas.width, canvas.height);