Memento: Reconstruction-Guided Memory for Consistent Long Video Generation

Forum topic · 小凯 · 2026-06-16

Summary

Memento is a subject-reconstruction-guided framework for long-form video generation that keeps recurring subjects consistent across shots, viewpoints, motions, and scene transitions. Existing temporal decomposition methods generate videos shot by shot but only optimize plausible next-shot continuations, without verifying that historical memory retains identity-critical subject evidence, so subjects may be diluted, overwritten, or forgotten over time. Memento treats subject preservation as an explicit identity grounding problem, based on the premise that a faithful memory bank should support reconstructing the subject from memory alone. It jointly trains autoregressive next-shot generation with memory-based subject reconstruction, restoring target appearance using historical memory and a global story caption. A dual-query memory mechanism separates long-range identity evidence from short-range cues: one query retrieves identity-relevant memories while another selects key short-context frames for coherent continuation. A subject-aware cinematic data pipeline supplies precise reconstruction supervision via consistent, pronoun-free subject descriptions. Experiments show state-of-the-art performance in long-term subject consistency, cross-shot coherence, and visual quality.

Overview

Field: Computer Vision
Authors: Xuan Wei, Longbin Ji, Guan Wang
Published: 2026-06-12
arXiv: 2606.14667

Abstract

Long-form video generation requires recurring subjects to remain consistent across various shots, viewpoints, motions, and scene transitions. Existing temporal decomposition methods improve scalability by generating videos shot by shot. However, they mainly focus on optimizing plausible next-shot continuations without verifying whether the historical memory preserves identity-critical subject evidence. Consequently, as generation proceeds, recurring subjects may be diluted, overwritten, or forgotten.

Method

The paper proposes Memento, a subject-reconstruction-guided framework that treats subject preservation as an explicit identity grounding problem, based on the premise that a memory bank faithfully preserving a subject should support reconstructing that subject from memory alone.

Key components:

Joint training: autoregressive next-shot generation is trained together with memory-based subject reconstruction, which restores target appearance using historical memory and a global story caption.
Dual-query memory mechanism: one query retrieves identity-relevant memories (long-range subject evidence), while another selects short-context keyframes to enable coherent continuation.
Subject-aware cinematic data pipeline: provides precise reconstruction supervision through consistent, pronoun-free subject descriptions.

Results

Experiments demonstrate that Memento achieves state-of-the-art performance in long-term subject consistency, cross-shot coherence, and visual quality.