Temporal Coherence in Video Watermark Removal
Video is not simply a sequence of isolated photographs—it is a continuous 3D spatio-temporal volume $I(x, y, t)$. When watermark removal algorithms treat each frame independently, minute variations between frames cause severe temporal flickering and luminance pulsing.
1. The Cause of Inter-Frame Flickering
When an AI video model (like Google Veo or Sora) generates footage, subtle video compression artifacts (H.264 macroblock boundary changes) and lighting fluctuations occur from frame to frame. If an automatic detector recalculates the watermark bounding box on every frame:
- Frame $k$: Bounding box detected at $(x=1820, y=1020)$.
- Frame $k+1$: Bounding box shifts by half a pixel to $(x=1820.5, y=1020.2)$.
This micro-jitter causes the cleaned patch boundary to vibrate rapidly at 24 to 60 Hz, creating a distracting buzzing flicker that ruins cinematic immersion.
2. The Static Coordinate Anchoring Strategy
AI generation engines place watermarks in static canvas coordinates. Recognizing this physical invariant, AURA ERASE executes a multi-frame consensus registration:
- Accumulation Phase: Sample the watermark corner across the first 15 frames ($t_0 dots t_{14}$).
- Temporal Median Consensus: Compute the median sub-pixel coordinate $(hat{x}, hat{y})$ across all candidate frames.
- Rigid Spatial Locking: Anchor the deblending mask permanently to $(hat{x}, hat{y})$ for every frame throughout the video duration.
3. Spatio-Temporal Boundary Blending
To eliminate edge seams where the cleaned region borders moving camera footage, a temporal exponential moving average (EMA) filter is applied to the boundary feather weights:
W_smoothed(x, y, t) = β × W(x, y, t) + (1 - β) × W_smoothed(x, y, t - 1) (β ≈ 0.85)
This guarantees that contrast transitions remain fluid and invisible, even during fast camera pans and dynamic lighting changes.
Experience Flicker-Free Video Cleaning
Upload your video clips to test temporal consensus anchoring in AURA ERASE.
Open Video Studio →