acceptodds
Under review as a conference paper at ICLR 2027

CERTIFIED CONSTRAINT LIFECYCLES FOR TRAJECTORY-PROXY VIDEO DIFFUSION CONTROL

Abstract

Constraint timing poses a dilemma in reverse denoising: later updates can violate previously satisfied constraints, while premature protection can hinder progress on unmet ones. To address this tension, we introduce Certified Constraint Lifecycles (CCL). CCL formulates inference-time diffusion control as an admission-preservation lifecycle under evolving protection obligations: qualifying observer-supported satisfaction on realized post-update states triggers irreversible gate-level admission, and the accumulated commitment state determines which protection relations and nonlinear endpoint checks apply to subsequent denoising updates. For differentiable trajectory proxies, CCL projects the sampler proposal onto the one-sided first-order cone induced by the exposed Jacobian and then applies a finite-step nonlinear guard. For successfully completed transitions, we prove guarded commitment invariance: every admitted proxy commitment remains satisfied. At a fixed state and proposal, with the exposed Jacobian evaluated exactly, we derive an exact pre-guard decomposition of the feasible row-space motion retained by one-sided protection but removed by equality protection. Controlled experiments show that CCL more reliably achieves and maintains trajectory-proxy satisfaction while reducing collateral deviation under matched final proxy satisfaction.

Primary area:
generative models
Keywords:
Video diffusion models, Inference-time control, Constrained sampling, Constraint lifecycles, Trajectory control

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_yhdUOpenReview ↗
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
  1. 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.
  2. 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
  1. 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).
  2. 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
  1. 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?
  2. 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
AuthorsOpenReview ↗

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).

ModelMethodSuccessFinal successTime per task (s)
ToraTora2529191.6
All-frame32361,014.3
State-local, 3 rounds3143518.9
CCL4547775.9
ReferenceMagicMotion–36753.1
Wan-MoveWan-Move2330756.5
State-local, 3 rounds3238993.1
State-local, 6 rounds32431,156.9
CCL41521,124.4
AuthorsOpenReview ↗

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.

Reviewer_JK37OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the challenges in video generation where later updates can violate previously satisfied constraints, and proposes Certified Constrained Lifecycles (CCL), a test-time diffusion control method that governs carry-forward protection during denoising. CCL achieves state-of-the-art (SoTA) performance in controlled experiments.

Critical strengths
  • Introduces the admission-preservation lifecycle, which transforms the qualification of post-update satisfaction into persistent commitments.

  • Establishes guarded commitment invariance for successfully completed transitions and provides rigorous theoretical proofs.

  • Demonstrates strong empirical effectiveness, showing clear improvements in acquisition and sustained coherence.

Critical weaknesses
  • High computational overhead: A primary weakness is the computational cost; the proposed method is approximately 39×39\times slower than the CogVideoX-2B base model and allocates about 5×5\times more GPU memory.
Questions for authors
  • Limited baseline comparison: Currently, the evaluation relies heavily on a single baseline (Tora). Incorporating stronger or more diverse baselines would significantly strengthen the evaluation.
1 comment
AuthorsOpenReview ↗

Response to Reviewer JK37

We thank the reviewer for the positive assessment and the constructive suggestions. References below are to the revised paper, where the object-level results are now Section 4.1 and Table 1 (formerly Section 4.3 and Table 4).

Weakness (cost): computational overhead.

We agree with the reviewer on the computational cost. Reducing it is the next step named in the revised conclusion, and Table 1 now reports each method's time per task. Most of the cost comes from the constraint readout and its Jacobian rather than from CCL's projection: in the CogVideoX-2b study that the reviewer cites, equality protection with the same observer and Jacobian costs nearly as much as CCL (Appendix F.2), and on Tora 79% of the added time goes to the soft readout of the tracked position and its Jacobian, against less than 1% for the projection (Appendix G.10). On the trajectory-control models the overhead is lower: CCL takes 1.49 times Wan-Move's time and 4.05 times Tora's, and on Tora peak memory grows by less than a quarter.

Question (baselines): comparison beyond Tora.

We agree, and the revision broadens the comparison beyond Tora (Table 1). On a second base model, Wan-Move, built on Wan2.1-I2V-14B, success rises from 23 to 41 of 64 tasks with CCL, much as it does on Tora (25 to 45), so the gain is not specific to one base model. A stronger trajectory-control method, MagicMotion, serves as a reference: at a generation time comparable to that of Tora+CCL, it reaches final success on 36 tasks, more than Tora alone (29) and fewer than Tora+CCL (47).

Reviewer_E4r9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Certified Constraint Lifecycles (CCL), a framework for inference-time trajectory control in video diffusion models. The proposed method introduces an admission-preservation mechanism in which constraints are admitted as persistent commitments after satisfying observer-supported conditions on realized denoising states. To preserve these commitments, CCL combines one-sided Jacobian-based projection with a nonlinear endpoint guard, aiming to maintain previously satisfied constraints while allowing compatible motion during subsequent denoising steps. The authors establish a conditional commitment-invariance result and derive a geometric decomposition comparing one-sided and equality-based protection. The method is evaluated through controlled proxy-constraint experiments, matched-satisfaction studies, and object-trajectory generation using CogVideoX and Tora. The experiments report improvements in constraint satisfaction, trajectory adherence, and collateral deviation, although the connection between the theoretical formulation and its different practical realizations remains insufficiently established.

Critical strengths

(1) A meaningful formulation of persistent constraint satisfaction during diffusion inference. The paper addresses an important limitation of inference-time constraint enforcement: satisfying a constraint at an intermediate denoising step does not guarantee that it remains satisfied after subsequent updates. The proposed admission-preservation mechanism provides a reasonable way to distinguish constraints that have already been acquired from those that remain unsatisfied. In particular, dynamically exposing protection relations according to the committed set is an appealing alternative to enforcing all constraints from the beginning. The combination of one-sided projection and nonlinear endpoint validation is also practically motivated, as it attempts to preserve existing commitments without unnecessarily restricting compatible denoising updates.

(2) Extensive empirical evaluation with useful component-level comparisons. The paper includes several controlled experiments examining admission timing, projection geometry, nonlinear guarding, and final-output deviation. The projection-guard factorial and comparisons between one-sided and equality protection help illustrate the different roles of the proposed components. Importantly, the external object-trajectory experiments evaluate generated motion using annotations independent of the online control observer, rather than relying exclusively on internal proxy satisfaction. The improvement in primary success from 25/64 for Tora to 45/64 for Tora+CCL is encouraging. The inclusion of computational measurements and additional failure analyses is also valuable for understanding the practical limitations of the framework.

Critical weaknesses

(1) The theoretical guarantees are narrower than the framework's broader motivation, and their connection to the implementation remains unclear. The central commitment-invariance result (Theorem 1) follows directly from the admission rule and nonlinear endpoint guard: accepted transitions preserve existing commitments, while new commitments are admitted only when their gates are satisfied. Although mathematically valid, this does not guarantee acquisition of previously unsatisfied constraints, nontrivial sampling progress, or meaningful object-level correctness under imperfect observers. Similarly, Corollary 2 establishes a fixed-state, pre-guard geometric advantage over equality protection, but not a corresponding advantage over complete denoising trajectories.

Moreover, Appendix E.2 reports finite-precision headroom behavior opposite to the ordering implied by exact projection geometry when dependent temporal relations are added. Although the authors acknowledge possible numerical-rank effects, the underlying mechanism remains unresolved. This raises questions about how well the idealized theoretical analysis explains the empirical improvements of the implemented controller.

(2) The experimental protocols differ substantially, making the central empirical contribution difficult to isolate. The matched-satisfaction experiments in Section 4.2 use initially feasible targets and a shared initial hold. While appropriate for evaluating collateral deviation under proxy satisfaction, they do not directly assess acquisition of independently specified trajectories. The large improvements over scalar rescaling should therefore be interpreted within this preservation-oriented setting.

The external Tora experiments in Section 4.3 address a more realistic task but use different observers, admission settings, and iterative corrections. In particular, they do not employ the structured position, velocity, and acceleration relations central to the earlier geometric analysis. This weakens the connection between the theoretical framework, proxy experiments, and practical object-level results.

Furthermore, Tora+CCL requires approximately four times the full generation time of Tora. Although the All-frame comparison is informative, it does not sufficiently isolate the benefit of persistent commitment management from additional iterative correction and computational effort under matched budgets. Despite extensive diagnostics, it therefore remains unclear which components primarily account for the reported improvements.

Questions for authors

(1) Relationship between theoretical geometry and numerical implementation. Appendix E.2 (Table 14) reports that the implementation-level headroom increases when velocity and acceleration relations are added, whereas the exact projection geometry implies the opposite ordering, HPVA≤HPV≤HPH_{\mathrm{PVA}} \leq H_{\mathrm{PV}} \leq H_{\mathrm{P}}. Although the authors acknowledge this discrepancy and discuss possible numerical-rank effects, the underlying mechanism remains unresolved.

Could the authors provide a more direct numerical analysis separating the equality- and one-sided-projection contributions, including the retained numerical ranks and relevant singular values? In particular, does the observed advantage persist when using a common factorized Jacobian or a numerically consistent projection operator? This clarification is important for understanding whether the theoretical decomposition explains the practical improvements or whether these improvements partly depend on implementation-specific numerical effects.

(2) Isolating the lifecycle contribution in the external object-trajectory experiments. The Tora experiments demonstrate encouraging improvements in object-level trajectory adherence. However, the practical implementation differs substantially from the structured projection framework developed earlier. In particular, Appendix G.3 uses DINOv2-based squared-position-error constraints and iterative relinearization rather than the structured position, velocity, and acceleration relations emphasized in the theoretical and proxy analyses.

Could the authors more clearly isolate the benefit of persistent commitment management from that of additional iterative correction? The All-frame ablation is informative about commitment-conditioned constraint exposure, but it does not fully establish how much of the improvement over the original Tora baseline is attributable specifically to the proposed lifecycle rather than to the additional correction mechanism. Given that Tora+CCL requires approximately 4.05 times the full generation time of Tora, a comparison under matched correction or computational budgets would be particularly valuable. This would help clarify which aspects of CCL are essential to its practical effectiveness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer E4r9 (Part 1/2)

We thank the reviewer for the detailed questions. To answer them, the revision adds a second base model (Wan-Move), a control that corrects without keeping commitments (State-local), a proxy-level benchmark with externally specified targets (Table 4), and an analysis of the numerical rank of equality protection (Appendix E.2–E.3). 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 endpoint guard is now the endpoint check, and Corollary 2 is Proposition 2.

W1 (theory): scope of Theorem 1 and Proposition 2.

We agree with both points: Theorem 1 follows directly from the admission rule and the endpoint check, and Proposition 2 is a fixed-state result. Theorem 1 states what a commitment guarantees, and the later results build on this guarantee. Because it rests only on these two rules, it holds regardless of how candidates are constructed, which is why it applies to the tracker-based controllers on Tora and Wan-Move as well, for the constraints as the tracker reads them.

Beyond the invariance, the paper examines whether commitments can be preserved while the remaining constraints are still acquired. At the proxy level, CCL completes 30 of 32 trajectories without stalling (Table 2), and Appendix A.5 gives sufficient local conditions under which the endpoint check accepts a positive step size. At the object level, under independent annotation, CCL raises success from 25 to 45 of 64 tasks on Tora and, in a new experiment, from 23 to 41 on Wan-Move (Table 1). For Proposition 2, the corresponding comparison over complete trajectories is experimental: CCL has lower deviation than equality protection on at least 83 of the 100 seeds under every metric (Section 4.3).

W1 (numerics) and Q1: the reversed ordering in Appendix E.2.

The reversed ordering comes from how equality protection sets its numerical rank, and it disappears when the three relation sets share one projector (Appendix E.2; Table 15, formerly Table 14). The table separates the two corrections. The one-sided correction grows as temporal differences are added, as the exact geometry predicts. The equality correction should not change, because temporal differences add no new directions; it grows only because its rank, set separately for each set, rises on average from 63 to 118 and 170, although at most 64 directions are independent. With one shared projector, the equality correction stays at 0.0130, the exact ordering HPVA≤HPV≤HPH_{\mathrm{PVA}} \le H_{\mathrm{PV}} \le H_{\mathrm{P}} holds, and the one-sided correction remains below it for all three sets.

RelationsOne-sided correctionEquality correction (mean rank), per-set rankEquality correction, one projector
Position (64)0.00400.0130 (63)0.0130
+ velocity (126)0.00640.0252 (118)0.0130
+ acceleration (186)0.00710.0349 (170)0.0130

CCL's advantage over equality protection does not come from this effect. Equality protection with position relations only is free of it: it has no dependent rows, and in exact arithmetic it gives the same update as with all relations. In a new comparison against this version, CCL still deviates less on 27 of the 32 seeds of Appendix E.3, by 0.0087 on average in pixel RMSE [0.0058, 0.0116] and 6.68 in latent distance [4.32, 9.03]. Adding temporal differences to both methods does not detectably change this margin (Appendix E.3). The object-level experiments and the comparison with scalar rescaling do not involve equality protection.

AuthorsOpenReview ↗

Response to Reviewer E4r9 (Part 2/2)

W2 (protocols): how the proxy-level and object-level studies connect.

The advantage over scalar rescaling is not confined to the preservation setting. We agree that the matched-satisfaction study measures collateral deviation rather than acquisition; acquisition of independently specified targets is evaluated in the new benchmark of Table 4, where targets are not satisfied at initialization and sampling uses the default schedule without the initial hold. With shared guidance, CCL succeeds on 73 of 128 tasks and scalar rescaling on 56 (+13.3 points [7.8, 19.5]).

The proxy-level and object-level studies share the lifecycle and the one-sided form of protection, and the lifecycle's effect carries over: protecting constraints before they are committed lowers completion from 30 to 17 of 32 proxy trajectories (Static-All, Table 2) and success from 45 to 32 of 64 Tora tasks (All-frame, Table 1). What differs is how the constraints are read. The tracker is not differentiable, so corrections go through a soft readout of the tracked position, in at most three linearized rounds until the tracker confirms every commitment. Because a single misread could trigger a commitment, admission uses K=2, which rules out such false admissions (Section 4.2). As the reviewer notes, each committed frame also contributes one row, its squared distance to the target, rather than the structured relations; this matches the tracker's constraint and keeps the Jacobian small (Appendix G.3). As noted in Part 1, Theorem 1 holds at both levels; Proposition 2 and the choice of relations are examined at the proxy level, and the object-level study tests the lifecycle.

W2 (isolation) and Q2: persistent commitments versus additional correction.

The revision adds the comparison the reviewer asks for, and it attributes at least half of CCL's gain in success to the lifecycle rather than to additional correction. Correction that is not restricted to committed frames recovers at most half of that gain, even with the same or a larger correction budget (Table 1):

ModelMethodConstrained framesSuccessFinal successTime per task (s)
ToraTora–2529191.6
All-frameall frames32361,014.3
State-local, 3 roundsframes failing now3143518.9
CCLcommitted frames4547775.9
Wan-MoveWan-Move–2330756.5
State-local, 3 roundsframes failing now3238993.1
State-local, 6 roundsframes failing now32431,156.9
CCLcommitted frames41521,124.4

On Tora, All-frame differs from CCL only in constraining all frames; it performs more correction (234 rounds against 220) and takes more time, yet succeeds on 13 fewer tasks than CCL (+20.3 points [9.4, 31.3]). On Wan-Move, the new State-local control keeps no commitments: at each update, it brings the frames that currently miss their targets within tolerance, using CCL's readout and solver (Appendix G.4 details how it differs from CCL). Even with twice CCL's round limit and slightly more time, it succeeds on nine fewer tasks (+14.1 points [3.1, 25.0]). What it lacks is persistence: it acquires the target in time as often as CCL (55 and 54 tasks), but 23 of these acquisitions do not end in success, against 13 for CCL, and doubling its round limit does not lower this number. The same pattern holds on Tora, where State-local acquires the target in time more often than CCL (54 against 49 tasks) yet succeeds on 31 against 45. These results suggest that the persistence of commitments is essential to CCL's practical effectiveness: without it, correction still acquires the target in time but does not maintain it until sampling is complete.

Thank you for the substantial revisions. I appreciate the authors' efforts to address my previous concerns, particularly through the additional numerical analysis and empirical comparisons.

However, my main reservations regarding the theoretical and methodological significance remain. Theorem 1 largely follows from the admission and endpoint-check rules, while Proposition 2 provides a fixed-state geometric result rather than a guarantee for complete denoising trajectories. Furthermore, the practical object-level implementation remains substantially different from the structured geometric formulation emphasized in the theoretical analysis.

Although the additional experiments strengthen the empirical evidence, I remain uncertain whether the overall contribution offers sufficient methodological novelty and theoretical depth to meet the standards expected at ICLR.

I therefore maintain my original rating of weak reject.

AuthorsOpenReview ↗

Follow-up to Reviewer E4r9: Methodological Novelty, Theory, and Implementation Unity

Thank you for your thoughtful follow-up. Three brief points on why CCL forms a unified methodological and theoretical contribution:

  • Methodological novelty (endogenous stateful control): Unlike memoryless guidance, prespecified schedules (e.g., SafeDiffuser), or classical active-set methods on a fixed optimization problem, CCL formulates inference-time control as a closed-loop state machine where realized denoising states dynamically expand the constraint problem subsequent updates solve. Our ablations show its three components form a minimal, non-trivial coupling: checking without projection frequently stalls (Check-only: 0/32 all-constraint success, 576/960 zero steps, Table 2); projection without checking regresses (25/32 seeds); protecting all frames upfront lowers success (All-frame: 32/64 vs. 45/64 on Tora; Static-All: 17/32 vs. 30/32); and memoryless correction loses acquired targets even at twice the correction-round limit (State-local 6-round: 32/64 vs. CCL's 41/64 on Wan-Move, at slightly higher runtime).
  • How the local theory drives protection and progress: Theorem 1, Prop. 2, Prop. 4, and Lemma 5 jointly explain why CCL preserves commitments while sustaining sampling progress. Enforcing Jtu=0J_t u = 0 (equality protection) removes the proposal's row-space component, dropping all-constraint success to 3/32 (vs. 30/32 for CCL, Table 8). Retaining the row-space cone component utretu_t^{\mathrm{ret}} (Jtutret≤0J_t u_t^{\mathrm{ret}} \le 0, Prop. 2) allows strict inward motion under Proposition 4's conditions, supplying the negative directional derivatives required for positive-step acceptance under Lemma 5 (30/32 all-constraint success with 0/960 zero steps, Table 2).
  • Implementation unity: At k=0k=0, the object-level QP (Eq. 26) reduces exactly to the cone-projection problem in Eq. 3, using the object-level squared-distance Jacobian. For the multistep samplers, up to three relinearizations construct candidate updates while retaining the native sampling schedule, with acceptance determined by the nonlinear endpoint check. Using one squared-distance row per committed frame (at most 49 rows on Tora vs. 288 structured PVA rows) directly matches the tracker's isotropic tolerance while keeping the Jacobian compact.

With the initial numerical-rank and budget-confounding concerns resolved, we hope this clarifies our core contribution and respectfully welcome any reconsideration of the soundness and overall scores.

Reviewer_2ASvOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes the CCL method, which facilitates decision-making by decomposing factors within the diffusion process; it also includes a mathematical derivation and proof. However, the overall writing is somewhat simplistic, and the proposed method is relatively basic. Furthermore, the experiments fail to sufficiently demonstrate the method's effectiveness, as there are no evaluations conducted on video generation models.

Critical strengths
  1. This paper proposes CCL, which facilitates decision-making by decomposing factors within the diffusion process;
  2. The paper is comprehensive and includes an appendix with supporting mathematical derivations.
Critical weaknesses
  1. The writing of this paper is poor. The title is written entirely in capital letters. The introduction fails to clearly articulate the core problem or the limitations of existing methods, leaving the advantages of CCL unconvincing. Furthermore, the flowcharts are rather crude, and the CCL framework itself is overly simplistic.
  2. The paper also lacks novelty. The proposed method is highly intuitive; while the appendix includes mathematical proofs, the underlying concept and approach are naive.
  3. The experiments are limited to "toy" cases. The method has not been tested on mainstream video generation models—such as Wan2.1 or Wan2.2—making it impossible to substantiate the effectiveness of the proposed approach.
Questions for authors

See the Weakness. The presentations and experiments are poor. Lack of the experiments of real video diffusion models.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2ASv

We thank the reviewer for the comments. We begin with the evaluation on video models, which the reviewer's question singles out. References below are to the revised paper.

W3 and Questions: evaluation on video generation models.

We agree that evaluation on mainstream video models is important. The submission evaluated CCL on two video diffusion models, and the revision adds a third, built on Wan2.1 as the reviewer suggests; the revised abstract now names all three. The first two are CogVideoX-2b, used in the controlled studies, and Tora, a trajectory-control model built on CogVideoX-5b, on which CCL raises success from 25 to 45 of 64 tasks under independent, blinded annotation (Section 4.3 of the submission, now Section 4.1). The third, Wan-Move, is a trajectory-control model built on Wan2.1-I2V-14B that generates 81 frames at 512×768; under the same annotation, CCL raises its success from 23 to 41 of 64 tasks (Table 1, Figure 5). In each case CCL is added at inference, without retraining the model.

W1 (writing): the introduction, the figures and the title.

We agree that the introduction should state the core problem and the limitations of existing methods directly, and the revised introduction now does. The problem is a dilemma: during denoising, a later update can undo a constraint that is already satisfied, yet protecting a constraint too early can hinder progress on those still unmet. Existing methods settle this in advance: guidance leaves satisfied constraints unprotected, while projection and barrier methods impose constraints from the start or on a prespecified schedule. CCL instead decides, from the realized denoising process, when each satisfied constraint becomes a commitment that later updates must preserve. The introduction then describes CCL's three steps in plain terms, and Figures 1 and 2 and their captions have been revised to use the same terms. The title appears in capitals because the ICLR style file sets it in small capitals.

W2 (novelty) and W1 (simplicity): what is new in CCL.

We agree that the idea is intuitive. What is new, relative to guidance (e.g., DPS) and to projection and barrier methods (e.g., SafeDiffuser), is to certify a constraint's satisfaction before protecting it and then to keep it: a constraint becomes a persistent commitment only after it has held for a prescribed number of consecutive updates, and every later update must preserve it. Simpler designs do not do as well: they recover at most half of CCL's gain in success over the base models (Table 1). Correcting whatever currently fails, without commitments (the new State-local control), reaches the target in time on more Tora tasks than CCL (54 against 49 of 64) but keeps it to the end on fewer (31 against 45), and in the new Wan-Move experiment it trails CCL by nine tasks even with its round limit doubled. Extending protection to all frames, rather than only committed frames, lowers success on Tora from 45 to 32.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer 2ASv,

Thank you again for your time and feedback. Following your suggestions, we have revised the presentation, clarified CCL’s contribution, and added experiments on Wan-Move, based on Wan2.1. We would be grateful if you could take another look at our response and the revised manuscript. If these clarifications and additional results address your main concerns, we would appreciate your reconsideration of the assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.