TL;DR of the paper, generated automatically.

New paper on keeping long video-to-video translations consistent in real time.

• Problem: real-time methods only see a fixed window of past frames.
• Method: matches input patches, then reuses earlier outputs from those same regions.
• Result: the paper says consistency improves across recent models while staying real-time.