Frozen History: When Temporal Losses Stop Accumulation in Learned Real-Time Supersampling

Key points

• Many methods use a temporal consistency loss to reduce flicker.1
• We show that with accurate registration, this loss can suppress accumulation: on static content, the network copies its warped history at almost no temporal cost.1
• We identify four registration errors that training can conceal: jitter-carrying motion vectors, one motion field for both resolution lattices, jittered depth, and an unregistered skip path.1
• With a small temporal weight and a normalised sample accumulator, a plain registered network improves the best published QRISP PSNR by 0.34 to 0.62 dB at $\times$2, $\times$3 and $\times$4 upscaling to 1080p.1

Sources (1)

• [1]Frozen History: When Temporal Losses Stop Accumulation in Learned Real-Time Supersampling
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 04:34 PM

  Learned real-time supersampling accumulates jittered low-resolution samples along motion vectors.

  Many methods use a temporal consistency loss to reduce flicker.

Extractive summary: sentences quoted from the sources.