Real-time supersampling techniques often work by accumulating jittered low-resolution samples along motion vectors, building up a higher-quality image over time. To keep the result stable, many learned methods add a temporal consistency loss that penalizes flicker between frames. This loss is generally treated as a safe way to smooth out rendering noise.
A new preprint on arXiv, "Frozen History: When Temporal Losses Stop Accumulation in Learned Real-Time Supersampling," challenges that assumption. The authors show that when motion registration is accurate, the temporal consistency loss can actually suppress the accumulation process itself. Instead of encouraging the history buffer to gather more samples, the loss freezes it, preventing further refinement.
The finding points to a subtle interaction between the loss objective and the accumulation mechanism. If the loss rewards consistency too strongly, the network may learn to stop updating the history, trading away potential quality for stability. The paper is a single source, so there is no independent comparison to note, but it highlights a failure mode worth considering when designing temporal losses for supersampling.