- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes FrameMorrow, a historical-frame selector for long-horizon video generation. Given the observed history, recent context, and currently available rollout condition, a lightweight causal Transformer autoregressively predicts a small set of prospective tokens. These tokens act as queries for ranking historical frames. During training, the selector learns from rankings derived from DINOv2 similarities between historical frames and realized future observations. Future observations are not required at inference. The selected frames are supplied to frozen generators through their existing reference or memory-conditioning mechanisms. The paper evaluates the approach across 11 generators spanning long-video generation, interactive generation, closed-source generation, and action-conditioned world models. It also provides selection baselines, design ablations, a diagnostic relating relevance scores to generation utility, and a user study.
Critical strengths
A practical selection interface with demonstrated reuse across generators.
FrameMorrow returns explicit historical frames rather than generator-specific internal states. The paper shares a selector checkpoint across compatible backbones within each condition modality and demonstrates integration with several frozen generators, including closed-source reference-conditioning interfaces. This provides meaningful evidence for the practical value of separating historical-frame selection from the generator.
Substantial evidence supporting the supervision strategy and design choices.
The paper compares selection strategies under matched historical candidates, frame budgets, and conditioning interfaces. Figure 6 examines future supervision and autoregressive versus parallel prospective queries.Appendix E.5 directly investigates whether future-grounded relevance corresponds to downstream generation utility: the predicted ranking achieves a Spearman correlation of 0.42, compared with 0.10 for current-context similarity. Although this diagnostic concerns individual-frame utility in a condition-free setting, it provides useful support for the proposed training proxy.
Critical weaknesses
-
The conceptual contribution relative to closely related work remains insufficiently distinguished.
As acknowledged in Section 2, Future Forcing constructs future-query proxies, PaFu-KV distills salience from a bidirectional teacher, and MemFlow retrieves history using the upcoming chunk prompt. Thus, anticipating future relevance is not by itself a sufficiently distinctive contribution. The generator-external, explicit-frame formulation is a meaningful difference, but the paper should more precisely establish what the learned prospective-token formulation adds beyond these related ideas. The existing ablations support its effectiveness within the proposed system, but do not fully resolve this question of conceptual differentiation. -
The evidence for the incremental advantage over strong selection baselines is not yet sufficiently conclusive.
In Table 5, FrameMorrow improves consistency over prompt matching by 0.95 points in the single-shot setting and 0.32 points in the multi-shot setting. These gains are encouraging, but their uncertainty is not reported. Furthermore, although the paper states that direct scoring uses the same inputs, its architecture, capacity, and training objective are not sufficiently specified to assess how cleanly this comparison isolates the prospective-token formulation. Improvements over native generators also include the benefit of adding historical references, as Appendix C.3 correctly notes, and therefore do not by themselves establish the additional value of the proposed selection mechanism.
Questions for authors
The following questions concern the interpretation of the current method and results.
-
Clarification of the contribution:
Relative to Future Forcing and PaFu-KV, what is the central methodological distinction of the prospective-token formulation beyond its generator-external, explicit-frame interface? Which existing results most directly support this distinction? -
Clarification of the direct-scoring control:
Table 5 states that direct scoring uses the same inputs. Please specify its architecture, trainable parameter count, and training objective. Does it use the same future-grounded ranking targets and optimization settings as FrameMorrow? -
Uncertainty in the central comparisons:
Could the authors quantify the uncertainty of the Table 5 differences against prompt matching and direct scoring, preferably using paired resampling of the existing per-case results? This would help determine how much weight to place on the reported incremental gains.