VISUAL MEMORY · LONG-VIDEO GENERATION
A relevant past can be wrong for the present.
We separate what memory retrieves from what a generator actually renders.
More recall can come with unwanted entities and outdated appearances.
Axi’s scarf has changed to a short cape. Retrieved history and the generated result keep the old scarf.
Outdated appearance
The history we want to remember.
Relevant to the past.
The old look is carried forward.
Wrong for the present.
Stills from earlier Track-B rollouts, shown in the paper’s Figure 1. One retrieved reference is displayed. These illustrate failure paths; they are not fixed-history causal interventions or samples from the main quantitative rerun.
Remembering an entity is only half the problem.
Memory must also respect what has changed.
Retrieval helps the right entities reappear. It also brings in entities that should be absent. A single consistency score misses this trade-off.
30 stories · 2,824 five-second segments per condition · seed 2026 · FLUX.2 + Wan2.2. Differences and 95% intervals use paired story-clustered bootstrap; they do not cover generation-seed variability.
| Condition | Recall ↑ | Precision ↑ | Avoid-OK ↑ |
|---|
A character can return after a long gap even with no visual memory, because the current prompt describes the character’s appearance.
On the same 62 long-gap character-return events, removing the appearance information sharply reduces no-memory presence. Long-gap success alone is insufficient evidence of visual-memory use.
Targeted prompt intervention; these are entity-presence rates, not proof of identity fidelity or a memory-capacity scaling law.
NO MEMORY · LONG-GAP CHARACTER PRESENCE
Same event cohort.
Different information in the current prompt.
Track A evaluates reference selection before generation. Track B evaluates the generated pixels. They answer different questions on different instances.
Choose references before the current real segment is revealed; observe the segment afterward.
No generator confound. The scoring gold is never an input to the method.
Generate causally from prompts and the system’s own history. Evaluate the entities and states that appear.
Four retrieval policies, no memory and a separate gold-ID Last-seen diagnostic control.
The 30-story evaluated cohort is part of a larger 50-story authoring pool. The two tracks are not matched instances, so Track-A scores cannot be directly correlated with Track-B outcomes.
Historical fidelity is useful only when the remembered state still agrees with the current story.
A needed entity may share its frame with a forbidden character, a look-alike or an obsolete appearance.
Separate gains from visual history from successes explained by the current prompt itself.
The main comparison measures complete retrieval policies on one shared generator, conditional on one rollout per story. Histories diverge, so it does not isolate the causal effect of replacing a single reference. The paper also reports second-judge and eight-story H3 sensitivity checks. Absolute rates depend on the judge, and the backend check has limited scope. Last-seen has privileged gold-ID access and variable reference counts; it is not a deployable method or a budget-matched upper bound. The 224-unit mechanism audit was performed by one author and does not estimate population failure frequencies.
arXiv:2610.35001 · Read the paper and download the PDF.
Scorers, method adapters, shared-generator testbed and reproduction guide.
03Text gold and prompts. Track B needs no source films; Track A requires separate video access.
THE PAPER
A Controlled Study of Visual Memory in Long-Video Generation
1 University of Science and Technology of China · 2 KlingAI Research · 3 Tsinghua University · 4 The Hong Kong University of Science and Technology · 5 City University of Hong Kong
@misc{chen2026vmembench,
title={More Memory, Better Videos? A Controlled Study of Visual Memory in Long-Video Generation},
author={Chen, Yuzhuo and Shi, Huafeng and Wang, Xinyu and Xu, Yulong and Wang, Yucheng and Hong, Haoqin and Zhang, Guoxin and Ma, Zehua and Zhang, Weiming},
year={2026}
}