Faithful text synthesis from low-resolution observations

Improve video diffusion models so that they can faithfully synthesize text that is barely legible in the initial viewpoint when the camera later approaches or zooms in.

Background

PPMem is evaluated on long-context camera-controlled video generation, where the model must preserve scene structure and visible content across substantial viewpoint changes. The reported failure case concerns fine text: text that is prominent and legible from the initial viewpoint can remain consistent, whereas text that is only barely legible becomes garbled as the camera approaches.

The authors explicitly identify faithful text synthesis from low-resolution observations as an unresolved challenge for video diffusion models. This problem concerns recovering or generating textual detail that was not adequately observed at the original input resolution, rather than merely maintaining geometric or semantic scene consistency.