Cause 1: frames generated with slight independence
Early methods restyled each frame as a separate image. Modern video models process many frames together, which links them, but sampling noise can still differ across frames.
Cause 2: unstable source signal
Auto-exposure shifts, compression blocks and sensor noise change pixel values frame to frame. The model interprets these as real changes.
Cause 3: ambiguity
Where the footage is ambiguous, such as foliage or crowds, the model picks different plausible interpretations over time.
What helps
- Clean, stable source
- Bolder, simpler styles
- Shorter segments
- Post-process temporal denoising