acceptodds
Under review as a conference paper at ICLR 2027

Measuring and Correcting Prior Transmission in Causal Video Diffusion

Abstract

A video that follows a requested trajectory may reflect effective control, motion that also appears in an unguided rollout from the same seed, or both. Yet evaluation based on target error alone does not reveal how much the controller improves on unguided generation or whether the remaining error reflects a predictable motion tendency. We introduce TrajectoryBench, a paired evaluation protocol that compares guided and unguided rollouts from matched conditioning and seeds, distinguishing final trajectory adherence from the improvement supplied by control. Across ten seeds on CausVid, controlled displacement is coupled to matched unguided displacement: roughly half of the unguided motion variation survives into the controlled output. We call the slope of this relationship prior transmission. The coupling persists after removing shared directional drift, survives replacement of the evaluation tracker, and appears on a second causal backbone, Causal Forcing. Predictable residual motion is not only an evaluation concern: it is a signal for improving control. We introduce Warp Forcing, a training-free trajectory controller for few-step causal video diffusion that steers objects block by block and adapts its commands to realized motion. Using the protocol's matched unguided rollout as a preview, drift compensation reduces mean path error relative to uncorrected Warp Forcing by 25% on CausVid and 38% on Causal Forcing across three seeds. A preview-free online estimator needs no extra rollout and adds no runtime: on Causal Forcing it matches the matched-preview correction, and on CausVid it recovers about one-third of the preview-based improvement. Measuring matched unguided motion thus yields both more informative evaluation and more accurate trajectory control.

Primary area:
generative models
Keywords:
motion guidance, training-free control, autoregressive video generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VUQLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper argues that target error alone conflates genuine controller contribution with motion the causal video generator would have produced anyway, and introduces TrajectoryBench, an 80-item paired protocol that compares each guided rollout with an unguided rollout under matched prompt, seed, and initial noise, reporting both error reduction and prior transmission, the slope between guided and matched unguided displacement along the requested axis. Across ten CausVid seeds the slope is about 0.52, survives seed-by-direction fixed effects, a wrong-seed derangement control, and a tracker swap, and reappears on Causal Forcing. The authors then propose Warp Forcing, a training-free blockwise controller that warps the clean estimate inside a tracked mask, rebuilds the KV cache from corrected frames, and replans each block; subtracting drift estimated from the matched unguided preview reduces mean path error by 25 percent on CausVid and 38 percent on CF, and a preview-free online estimator matches the gain on CF while recovering about one third of it on CausVid on a single seed.

Critical strengths
  1. Seed-matched paired evaluation is a simple and sound way to separate control contribution from motion the model would produce anyway.
  2. The transmission analysis is statistically careful, with object and seed clustering, wrong-seed derangement, seed-by-direction fixed effects, and an independent tracker check.
  3. The measurement directly motivates the correction, and the derangement control shows that item-specific drift, not only generic drift rejection, drives part of the gain.
  4. Ablations on cache reconstruction, replanning, and injection step give a coherent picture of which components matter.
Critical weaknesses
  1. The controller is not practically competitive, as native bidirectional baselines reach 23 to 29px path error while corrected WF-r still sits near 91 to 93px on CF and 35.7px on CausVid.
  2. Absolute adherence on CF remains poor, with a large fraction of leftward items still moving opposite the request after correction. Drift compensation is standard feedforward disturbance rejection and the paired unguided baseline is an obvious control, so conceptual novelty is limited relative to the empirical effort.
  3. Evaluation is narrow, covering template prompts, centroid-only rigid translation, mostly cardinal 150px moves, no user study, and quality judged mainly by global DOVER and sharpness.
  4. Prior transmission is explicitly a predictive statistic rather than a causal quantity, so its value as a benchmark metric is interpretive rather than diagnostic.
  5. Matched-preview correction measurably degrades CF background fidelity, with background LPIPS rising from 0.258 to 0.308 and PSNR dropping from 17.98 to 16.87dB.
  6. The preview-free estimator comparison rests on a single seed, so the practical claim that online estimation matches matched-preview correction on CF is not established at the same evidence level as the main result.
  7. The paper is hard to follow because results are spread over many campaigns with different splits, seeds, and item counts, and key claims require cross-referencing multiple appendix tables.
Questions for authors
  1. Can the online estimator comparison be repeated over the full three seeds on both backbones, with object-clustered intervals against matched preview.
  2. Why does WF-r on CF still perform worse than a stationary object on many leftward items, and would forward-predicting the mask support with the estimated drift address this.
  3. Since the matched preview and the evaluation both use CoTracker3 on the same base clip, have you computed the preview drift with a different tracker to rule out tracker-specific shared bias.
  4. Can you show a case where prior transmission changes a design decision or controller ranking that path error and error reduction alone would not.
  5. How sensitive are the slope and the correction gain to prompt richness, given that the template prompt deliberately leaves motion underspecified.
  6. Can you report CF results separately by direction, since the rightward bias appears to dominate the aggregate numbers and may mask per-direction behavior.
Reviewer_jUdJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces TrajectoryBench and Warp Forcing to measure and correct the effect of unguided motion on controlled video generation. Unlike evaluation based only on how closely a generated video follows the given trajectory, this protocal asks how much the controller improves over a matched unguided generation. The authors use motion predicted from the unguided output as a correction signal. Specifically, the authors pair guided/unguided videos, then track the target object's trajectory. They measure the reduction in path error and the regression slope between guided and unguided displacement. Warp Forcing shifts the object region selected by a mask within the predicted clean latent, adds noise to the result, and uses the remaining denoising steps to refine the edit. The authors show that correcting for unguided motion improves trajectory tracking on CausVid and Causal Forcing.

Critical strengths
  1. Useful empirical observation. The paper shows that motion in a matched unguided video can help predict the motion that remains after applying trajectory control.
Critical weaknesses
  1. Unclear contribution. TrajectoryBench provides an additional measurement, but its mean error-reduction metric does not change controller rankings when all methods share the same unguided baseline. Warp Forcing combines familiar latent editing, re-denoising, and feedback correction, leaving its methodological novelty unclear. There are many prior works on warping-based method. The paper should clarify its technical contribution.

  2. Poor readability and presentation. The paper would benefit from substantial polishing. Repeated explanations and secondary details (e.g., CI, ) make the main contribution difficult to identify. Figure 1 is hard to follow, with small annotations and no clear difference between methods. As noted in the AI use statement, LLMs seem to have been used for writing, so the paper still needs a lot of polishing.

Questions for authors

Figure 1 and its results are difficult to understand. Could the authors provide qualitative comparisons with video?

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

This paper pressents two coupled contributions:

TrajectoryBench — a paired evaluation protocol. Run guided and unguided rollouts under identical prompt/seed/initial noise, use the unguided rollout as a reference, and separate "final trajectory adherence" from "the improvement actually supplied by the controller." Defines prior transmission as the slope β in y = α + βx, where y and x are guided and unguided net displacement projected onto the requested axis. On CausVid with 10 seeds × 20 objects × 4 directions (800 obs), β = 0.523 — roughly half of the unguided motion variation survives into the controlled output. Warp Forcing — a training-free trajectory controller for few-step causal video diffusion. At step k*, apply a mask-scoped warp to the clean estimate x̂₀ → re-noise using the scheduler's own forward rule → finish the remaining steps → rebuild the KV cache from the corrected content → replan each block from online SAM2 tracking with a shrinking horizon. Then subtract the measured predictable drift as a feedforward disturbance: rate_b = (c_tot − c_b)/B − g·δ_prior, cutting path error by 25% (CausVid) and 38% (CF).

Critical strengths

1.The measurement insight has value Tora (TrajError) and DragStream (REVEL, ObjMC) score only against the requested path. The contamination source — "an underspecified prompt plus a fixed seed already produces motion toward the target" — has genuinely been overlooked by the subfield. This is a legitimate evaluation-hygiene contribution

2.Several diagnostics are excellent ① Zero-flow sham: the full control stack with an identically zero flow field gives slope 1.00, r = 1.0, validating the pairing. ② Candidate regeneration: a block commanded to move 42.4px realizes 106.4/106.9/101.4px across three draws — strong evidence the disturbance is not sampling noise. ③ CF's rightward drift is as large or larger under the bare prompt "A {object}." (+197 vs +181px), ruling out the prompt and attributing it to checkpoint rollout dynamics. ④ Decomposing the disturbance into parallel/orthogonal components explains why netCos saturates at 0.97 while 53px of path error remains

Critical weaknesses

1.The primary metric is debatable. trajPathErr uses a constant-rate reference ramp, penalizing clips that follow the right path with different timing; the authors' own DTW cross-check shows 53.5→33.3 and 34.5→19.3, i.e. a large share of reported "path error" is timing, not geometry. FDA is systematically biased toward 0.5 on near-stationary clips by the 10⁻⁶ convention. netCos saturates at 0.97. Reviewers will ask why not ObjMC/TrajError as used by Tora/DragStream — which would also fix comparability.

2.The method actually fails on CF, and percentages hide it. At a 150px request, a stationary object scores trajPathErr = 75px. On CF: unguided 203.6 → WF-r 158.8 → corrected pooled 93.1px. That is, with the full method applied, CF is still on average worse than not moving the object at all. The paper reports only "the fraction exceeding the stationary-object error falls from 75% to 30%" and never takes the knife to its own mean. The CausVid side (91.7→53.5→35.7) is sound; the CF "38% improvement" is improvement over a very bad baseline.

3.No credible SOTA comparison Every win is over a self-made transplant. Table 3: native FreeTraj/VideoCrafter2 = 23.0px, SG-I2V/SVD = 29.0px; WF-r = 53.5 (CausVid) / 158.8 (CF). The only places WF-r wins (Tables 2/19) are the authors' own ports of other mechanisms into CausVid, using released defaults with "no dedicated causal tuning" (FreeTraj 114.9, DragStream 87.2, QWERTY 72.4). DragStream is precisely the real same-lane training-free streaming competitor, yet it is never run in its native setting or on REVEL. This is a textbook weak-baseline objection.

  1. The "transmission and accuracy are distinct properties" claim is weakly supported. The overall trend in Table 20 is actually low transmission ↔ low error (native SG-I2V −0.10/29.0px; native FreeTraj +0.03/23.0px; raw-latent 25it 0.18/40.6px). The dissociation rests on basically one or two cells (FreeTraj transplant: slope ≈ 0, worst error). The more dangerous inference: if β ≈ 0 for strong controllers, prior transmission may be a diagnostic of weak control rather than a universal evaluation flaw — which directly undermines the protocol's importance.
Questions for authors

please see the weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.