- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2(+1)
- Soundness
- 2
Read review
Paper summary
CCL irreversibly admits trajectory constraints once satisfied for K steps, then projects later updates onto a one-sided linearized cone and line-searches to keep admitted constraints satisfied. It proves that admitted constraints stay satisfied and that one-sided projection deviates less than equality projection. Validation uses CogVideoX-2b protocols and Tora on 64 tasks with human-verified annotations.
Critical strengths
- On Tora, scored independently of the controller, success rises from 25/64 to 45/64. This is the most convincing evidence. The gain holds for both linear and curved paths and includes better acquisition, not only better preservation.
- The statistics (preregistration, paired seeds, Holm correction) are careful. This makes the within-protocol comparisons, such as one-sided vs equality protection, reasonably reliable.
Critical weaknesses
- The CogVideoX setup seems artificial. App. B.2 states that "Targets are deterministically blended toward the observer readout" so that "max_k c_k(z_init) ≤ 0.8ϵ" (p. 17, l. 904–907), begins with "a six-transition shared hold" and scales proposals by "γ = 0.5" (l. 911–917). App. F.7 reports "Dynamic degree is zero for every arm" (p. 37, l. 1964). The reported "82.21% to 89.38%" reduction (Sec. 4.2, p. 7, l. 373) is against uniform step shrinkage, and there is no comparison to existing trajectory-control or plain guidance methods. Under these conditions the deviation numbers mainly show that shrinking the whole step is costly, not that CCL controls motion better than practical alternatives. On Tora, the comparison is also limited to with and without CCL plus an All-frame variant (Table 4, p. 8, l. 397).
- The contribution "we prove guarded commitment invariance" (Sec. 1, p. 2, l. 089) largely follows from the construction of Theorem 1 (p. 5, l. 240); App. A.1 itself notes "This invariant concerns preservation after admission; it neither establishes ..." (p. 13, l. 672). Corollary 2 rests on "Moreau's decomposition" (App. A.3, p. 14, l. 717). Runtime is also high: "1243.72 s, or 38.95 times" the baseline (App. F.2, p. 33, l. 1781). On Tora the overhead is about 4× ("775.9 seconds per task on average, compared with 191.6", App. G.9, p. 45, l. 2400), and the main text does not discuss cost. The underlying algorithm (a projection, a line search and a dwell counter) is fairly simple but hard to see through the terminology.
Questions for authors
- Table 2 (p. 7, l. 322–329) compares only internal handlers such as "Static-All Restoration". How does CCL compare at matched compute with gradient guidance on the same proxy and with an existing trajectory-control method?
- App. B.2 (p. 17, l. 904–917): what are the results under the default sampler, without "γ = 0.5", and with targets not satisfied at initialization?
3 comments
Response to Reviewer yhdU (Part 2/2)
W2 (theory): Theorem 1 and Corollary 2 follow from the construction.
We agree: Theorem 1 follows directly from the admission rule and the endpoint check, and Proposition 2 is a direct application of Moreau's decomposition. Theorem 1 states what a commitment guarantees; the later results build on this guarantee. Because it rests only on these two rules, it holds at every realized state regardless of how candidates are constructed, which is why the same guarantee carries over to the tracker-based controllers on Tora and in the new Wan-Move experiment, for the constraints as the tracker reads them. We also agree that preservation alone is not sufficient: Check-only satisfies Theorem 1 too, yet stalls on 576 of 960 updates and completes no trajectory. Beyond the invariance, the paper therefore examines whether commitments can be preserved while the remaining constraints are still acquired: Appendix A.5 gives sufficient local conditions under which the endpoint check accepts a positive step size, and CCL completes 30 of 32 trajectories without stalling (Table 2). Proposition 2 identifies, at a fixed state, the component of the proposal that one-sided projection keeps and equality protection removes.
W2 (cost): runtime overhead and its discussion in the main text.
We agree that the main text should report the cost, and the revision now does: Table 1 lists each method's time per task, and the conclusion names reducing it as the next step. In the new Wan-Move experiment the overhead is smaller than on Tora: CCL takes 1.49 times the base model's time. Most of the cost lies in reading the constraints rather than in CCL's projection: of the time added on Tora, 79% goes to the soft readout of the tracked position and its Jacobian, and less than 1% to the projection (Appendix G.10). The same holds in the proxy study, where equality protection with the same observer and Jacobian costs nearly as much as CCL (Appendix F.2). Nor does the extra time produce the gain: as shown in Part 1, correction with more time does not reproduce it.
W2 (presentation): the algorithm and its terminology.
We agree that the algorithm should be easy to see, and the revision restates it in plain terms: the introduction now gives each of the three steps in one sentence, and the finite-step guard, for example, is now the endpoint check. We see the simplicity itself as a strength: CCL adds three steps to an existing sampler without retraining. The core idea is how the counter governs the other two steps: a constraint enters the set that the projection and line search protect only after it has been satisfied at K consecutive updates, and it remains there until sampling ends. The protected set is therefore not fixed in advance but grows with the realized denoising process.
Thanks for the clarification. Most of my major concerns have been addressed, so I’ve raised my score to 3.
Response to Reviewer yhdU (Part 1/2)
We thank the reviewer for the careful reading and for finding the Tora result convincing. To answer the questions, the revision adds a second base model (Wan-Move), a control that corrects without keeping commitments (State-local), MagicMotion as a reference, and two benchmarks under the default sampler with targets not satisfied at initialization (Table 4). References below are to the revised paper, which reorders Section 4: object trajectories are now Section 4.1 and Table 1 (formerly 4.3 and Table 4), staggered admission Section 4.2 and Table 2 (formerly 4.1 and Table 1), and matched satisfaction Section 4.3 (formerly 4.2). The constraint-handler comparison (formerly Table 2) is removed, Guard-only is now Check-only, and Corollary 2 is Proposition 2. In the appendix, the settings formerly in B.2 are now in B.6 and C.2, F.7 is now F.5, and G.9 is now G.10.
W1 (setup) and Q2: the CogVideoX setting; results under the default sampler.
CCL's advantage over scalar rescaling holds under the default sampler as well, and there CCL keeps the video's own motion. We agree that the matched-satisfaction setting is a controlled one: it isolates how protection is applied. The 82%–89% reduction is relative to whole-step shrinkage; relative to equality protection, which differs from CCL only in how the relations are enforced, the reduction is 20.86% to 29.64% (Table 3).
To answer the reviewer's second question, Table 4 adds two benchmarks under the default sampler, without the hold or γ = 0.5, whose targets are not satisfied at initialization. On 128 external targets fixed before generation, with the same guidance for all methods, CCL succeeds on 73 tasks and scalar rescaling on 56 (+13.3 points [7.8, 19.5]). Equality protection performs similarly there (71 tasks; +1.6 points [−2.3, 6.3]) but not on the second benchmark, whose targets are derived from unconstrained readouts: CCL succeeds on 28 of 40 conditions, equality protection on 17, and scalar rescaling on 20. The other two methods succeed less often there mainly because they suppress the video's own motion: 32 of CCL's 40 videos pass the VBench dynamic-degree criterion, against at most 24 for either of them. On the external targets, 127 of CCL's 128 videos pass it, as many as with unconstrained sampling (Appendix D.8). The advantage is therefore not confined to the controlled setting.
W1 (baselines) and Q1: comparison with guidance and trajectory-control methods at matched compute.
CCL is added on top of these methods rather than replacing them, so we compare each with and without CCL. With gradient guidance on the same proxy held fixed, adding CCL raises success from 66 to 73 of the 128 external-target tasks (+5.5 points [0.0, 10.9]; Table 4). On trajectory-control models, it raises success from 25 to 45 of 64 tasks on Tora and, in a new experiment, from 23 to 41 on Wan-Move. At a comparable generation time (775.9 s against 753.1 s), Tora+CCL also reaches 47 final successes against 36 for the newly added MagicMotion (+17.2 points [−3.1, 37.5]), a difference that is not statistically resolved.
At the object level, the gain is not bought with compute: correction that spends more time without being restricted to committed frames recovers at most half of it in success. All-frame succeeds on 32 Tora tasks against 45, and the new State-local control, even with its round limit doubled, on 32 Wan-Move tasks against 41 (+14.1 points [3.1, 25.0]). Correction at the current state acquires the targets as often as CCL but keeps fewer of them: with six rounds, State-local acquires the full target in time on 55 Wan-Move tasks, against 54 for CCL, yet 23 of these acquisitions do not end in success, against 13 (Table 1).