UniTemp: One 1.3B Model Generates Video in Any Temporal Order via Bidirectional Distillation

Researchers have introduced UniTemp, a single 1.3B-parameter autoregressive video model capable of generating video in any temporal direction—forward, backward, or interpolated between two given frames—using just four denoising steps. The model matches dedicated single-direction baselines in each mode without per-direction specialization, and its compact design supports stable generation of videos up to 120 seconds long.

The unified temporal control enables capabilities that forward-only models cannot easily achieve: seamless looping video, in-shot scene transitions between two distinct environments, and non-chronological multi-scene storytelling. In the latter workflow, key scene blocks are generated first, extended bidirectionally, and then connected via generated transitions—including a closing loop from the final scene back to the first—demonstrating a more flexible pipeline than conventional chronological generation.