acceptodds
Under review as a conference paper at ICLR 2027

PhysReflex: Counterfactual Cyclic Refinement for Physics-Consistent Video Generation

Abstract

Modern video generators achieve impressive visual fidelity, yet still struggle to faithfully capture the physical event specified by the prompt. This failure is particularly subtle when different event outcomes produce similar-looking motion, making current-state refinement potentially insufficient for distinguishing whether the intended dynamics are being formed. Existing methods address this challenge through physics-aware training, explicit physical priors, or inference-time guidance and refinement. However, training-free inference methods typically correct generation based on the current sampling state, without explicitly examining how different physical evolutions unfold from that state. We introduce PhysReflex, a training-free framework for counterfactual cyclic refinement. At selected sampling steps, PhysReflex briefly advances the current latent under the target condition, then traces it back under both the target and a counterfactual condition. The discrepancy between the returned states provides an evolution-aware signal for correcting the target sampling trajectory. We further adapt the number of refinement cycles according to changes in the predicted clean-video latent, avoiding unnecessary or excessive correction. Experiments on VideoPhy2 and PhyWorldBench demonstrate consistent improvements over baselines across comprehensive evaluations of physical plausibility, motion consistency, and semantic alignment.

Primary area:
generative models
Keywords:
Video Generation, Physics-Aware Video Generation, Counterfactual Refinement, Training-Free Refinement, Adaptive Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zTFQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysReflex is a training-free sampler for improving prompt-faithful physical events. At sparse anchors, it steps forward under the target prompt, integrates back under both the target and a physically valid counterfactual prompt, and adds the normalized difference to the latent. An adaptive gate decides whether to repeat. On Wan2.2-A14B, it improves VideoPhy2 semantic alignment by +8.3 / +7.5 points at ~1.5× cost, and also improves on a PhyWorldBench subset.

Critical strengths
  • The idea is simple and fits the problem. Probing how two event conditions respond after a short look-ahead along the target trajectory, with the target's own return as a matched reference, is a clean construction.

  • The baselines are the right mechanism controls. Direct counterfactual guidance and SDG at matched call budgets are exactly what you need to show that the cycle matters. The direction of improvement holds across two judges and two benchmarks.

Critical weaknesses
  • The experiments don't directly test the central claim. The paper is motivated by event-attribute fidelity, but nothing in the evaluation directly measures it. VideoPhy2 SA and PhyWorldBench Content are frame-sampled VLM judgments of general prompt adherence. The authors themselves note (App. L) that a frame-based judge may not reliably measure deceleration rate, and they leave trajectory-sensitive tests of braking rate.

  • The counterfactual prompt c− is an extra input, and the paper does not say who or what wrote it for the VideoPhy2 and PhyWorldBench prompts. App. D lists only the pairing rules, and the only examples given are illustrative (braking, bounce count, impact).

  • The mechanism may reduce to a known baseline. For m = 1, the anchor cancels (App. B, Eq. 12-14) and the update is simply a direct counterfactual contrast evaluated at the one-step-denoised state y and σ_{i+1}. The h² "state-response" term in Eq. 9 is just the effect of evaluating there instead of at the anchor; genuinely path-dependent terms appear only for m > 1, yet m is never reported or ablated. Since the Direct Counterfactual control is evaluated at the anchor (Eq. 7, Sec. 4.2), a "direct contrast at y, no return" baseline and an m sweep are needed to show that the return path adds anything.

Questions for authors
  1. Can you evaluate the central claim directly? For example, for a set of prompts with a measurable attribute (braking rate, bounce count, impact strength), generate from the same seed with (c+, c−) and with the swapped pair (c−, c+). Then check whether the outputs actually separate along the attribute.
  2. How did you write c− for the benchmark prompts? How many pairs were rejected under the App. D rules?
  3. Value of m and the return path. What value of m is used in the reported experiments?
  4. Please report the values of the anchor set A, η, τ and B, and the distribution of accepted cycles per sample. How does PhysReflex compare with a fixed number of cycles and no gate?
Reviewer_s9bFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PhysReflex, a training-free inference-time correction method for frozen text-to-video diffusion models. It targets cases where generated motion is physically plausible but does not realize the specific physical event requested by the prompt. PhysReflex uses a matched counterfactual prompt pair that preserves the scene and changes one physically admissible event attribute. At sparse denoising anchors, it moves briefly along the target-conditioned trajectory, returns from the shared endpoint under both target and counterfactual conditions, and applies the return difference to correct the target state. Probe paths are discarded, and only the target trajectory continues. An adaptive gate allocates an additional probe when the target clean prediction remains responsive, subject to a reserved budget. Experiments show the largest gains in semantic alignment by PhysReflex, with smaller improvements in physical commonsense and joint success.

Critical strengths
  1. By introducing and using a counterfactual condition, the paper handles the negative condition in a way different from CFG, achieving generation refinement.
  2. The proposed method is training-free; it only requires modifying the model inference process to achieve refinement.
Critical weaknesses
  1. As a general method, the paper only conducts experiments and evaluate results on the Wan2.2-A14B model, which is insufficient.
  2. The paper does not describe in detail how the anchor set is selected, and it does not provide sensitivity experiments on the method’s dependence on anchor set selection.
  3. As a refinement rather than a prompt-alignment enhancement approach, practical users may need to first generate with the original prompt to identify where reinforcement is needed, and then find a suitable counterfactual condition themselves, which itself could be hard to identify.
  4. For counterfactual condition construction, the paper does not provide a general method. The paper does not provide a suitable method for this process, nor does it provide sensitivity experiments.
Questions for authors
  1. In App. D, the authors propose a counterfactual construction protocol and mention that counterfactuals that do not conform to the protocol will be rejected. Did the authors actually implement this protocol? How are counterfactuals identified as valid or invalid, and how is correction performed when they are invalid?
  2. If we generalize the counterfactual condition to a general negative prompt (i.e., the logic of CFG, in which case this paper and CFG would differ only in computation), why is this not applicable? Why is the proposed method only applicable to physical counterfactual conditions? Why is it necessary to restrict it to the protocol in App. D?
Reviewer_2tskOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces PhysReflex, a training-free sampling method for generating videos which follow physical event in prompt. Unlike direct conterfactual guidance evaluated at the current noisy state, it compares conditional responses after a short lookahead. Specifically, it creates a counterfactual prompt, and at selected denoising steps, it takes a few inference steps, then runs backward to the original noise level. It subtracts the returned target latent from the returned counterfactual latent and normalizes the difference to update the main latent. Additionally, the authors propose an adaptive gating method based on how much the predicted clean latent changes after correction. It allows another cycle when this change exceeds a threshold and enough computation budget remains.

Critical strengths
  1. The paper tackles a practical problem: prompt following in physically plausible video generation.
Critical weaknesses
  1. Limited experimental support. Table 2 Gemini-Direct counterfactual: Joint shows higher score than Physics. Additionally, the evaluation only covers one backbone and two simple benchmarks, including only a subset of PhyWorldBench. The qualitative results are difficult to interpret, and no videos are included in the supplementary materials. These inconsistencies in experimental results need clarification. Additionally, given prior findings that overall motion is largely determined early in denoising (like in Self-Refining video sampling), the effectiveness of the proposed lookahead should be examined more carefully across different denoising stages.

  2. Limited novelty and unclear source of gains. Forward (lookahead) backward refinement is a common method in generative models. Given the existing model’s difficulty in capturing small difference in prompts, it is unclear whether the counterfactual response reliably reflects the intended event difference. Additionally, the current comparisons do not clearly separate the gains from iterative refinement and those from lookahead. Ablation studies are needed to establish the contribution of each component.

Questions for authors
  1. Could the authors provide playable videos? Does proposed method cause over-saturation?

  2. See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.