Memento: Reconstruct to Remember for Consistent Long Video Generation

Researchers have proposed two new methods for generating consistent long videos and cardiac sequences, improving upon existing techniques in terms of subject consistency and visual quality.

The first method, called Memento, is a framework for consistent long video generation by preserving subject identity across shots and scene transitions^[1]. Memento treats subject preservation as an explicit identity grounding problem and jointly trains autoregressive next-shot generation with memory-based subject reconstruction. It introduces a dual-query memory mechanism to disentangle long-range subject evidence from short-range cues. According to the researchers, Memento achieves state-of-the-art performance in long-term subject consistency, cross-shot coherence, and visual quality. A second, separate research effort proposes a generative method for synthesizing temporally coherent and anatomically consistent cardiac sequences. This method decouples cardiac spatial structure from temporal motion and generates diverse sequences with high temporal coherence and strong fidelity to input prompts^[2]. The model achieves a FID of 31.68 for image realism and a CLIP score of 31.04 for text-image alignment.