acceptodds
Under review as a conference paper at ICLR 2027

Causal-Role Erasure for Text-to-Video Diffusion Models

Abstract

Video object removal is an important capability with broad applications in controllable generation and content editing. However, erasing an object does not always erase its physical causal effects: a droplet may vanish while its splash remains. Existing text-to-video concept-erasure methods target named concepts rather than their causal roles, leaving this conditional removal without an explicit formulation or training strategy. We formulate causal-role erasure as suppressing a source and its effects when it causes an event while preserving the same object's noncausal uses, and introduce Source-slot Randomized Counterfactual Distillation (SRCD) to address it. Specifically, we construct independently generated and screened source-free target videos that retain the receiver and scene without the source or its effects, then pair each target with event prompts that vary only the source identity. Across seven physical mechanisms, SRCD obtains a higher average graded erasure score than an otherwise matched control, but the corrected difference is not statistically significant; paired-role tests show no clear additional benefit, and human review does not confirm the automated comparison.

Primary area:
generative models
Keywords:
text-to-video generation, diffusion models, concept erasure, counterfactual distillation, role-conditioned generalization

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

This paper studies causal-role erasure in text-to-video diffusion models: removing an event source and its physical effects while preserving the same object when it appears in a noncausal role. The proposed method, Source-slot Randomized Counterfactual Distillation (SRCD), varies the source identity in training prompts while keeping independently generated source-free target videos fixed. It trains mechanism-specific LoRA adapters through flow matching, teacher distillation, and preservation updates. The paper evaluates seven physical mechanisms. The setting is interesting, particularly because inference requires only a text prompt and the corresponding adapter. However, the evidence for a reliable improvement over the matched control remains limited.

Critical strengths
  1. The causal-role erasure setting is novel.
  2. The controlled single-factor design is rigorous (Holm correction, non-inferiority margins).
  3. Negative results are reported candidly.
Critical weaknesses
  1. Each adapter is trained on only 100 erasure pairs. Is this sufficient? Does the model learn the scenes in the training set or the prompt descriptions, or does it truly generalize? Does SRCD’s advantage still hold within the trained mechanisms for unseen object–receiver combinations, substantially different scene layouts, and free-form prompts?

  2. What does the learning curve with respect to data size look like? The authors could compare several training-set sizes while controlling the optimization budget and repeating seeds, and check whether erasure, role preservation, and generalization improve with more data.

  3. As described in the paper, effect-aware removal methods can handle such effects using an observed video, masks, or region localization. Their limitation is not that they cannot remove these effects; it is that they cannot be used directly through the text-only generator interface defined here. Nevertheless, the paper provides no direct experimental comparison with these methods. Moreover, the training prompts already describe the physical consequences, and the source-free targets are themselves generated by the base model from source-free prompts. Could a carefully designed prompt achieve the same result directly? I suggest adding such a naive baseline.

  4. Figure 2 does not clearly show the full computation path: how the target video is encoded and noised, which text the teacher and student each receive, which parameters are updated, and which modules remain at inference.

  5. Figure 3 shows clear failures in which the source and its effects are not fully erased. Consider these alongside the complete-absence metrics, the statistical tests, and the human evaluation, the current evidence neither establishes a reliable advantage of SRCD over the matched control nor shows that it can consistently accomplish the defined task.

Questions for authors
  1. Please clarify whether the target videos were screened by humans or by an automatic evaluator. Please also report the residual error rate after screening, along with independent human spot-check results. The source-free prompt does not depend on the source, and each mechanism has only 12 receivers and two wording styles. How many distinct scene conditions do the 178 targets per mechanism actually contain? What is the rationale for using only 100 pairs per adapter?

  2. When a scene contains multiple mechanisms or multiple active event sources, which adapter should be used? Can the method remove all downstream consequences of a specified source while preserving similar effects produced by other causes? Would this setting require retraining?

  3. Compared with flow-matching fine-tuning alone on the same target videos, how much does teacher distillation contribute? Can this gain, and the gain from source replacement, be reproduced across multiple training seeds?

1 comment
AuthorsOpenReview ↗

Response to Reviewer wsPZ

Thank you for recognizing the task formulation and controlled comparison. We would like to clarify two points:

  1. The 100 training pairs per adapter are the setting used in this study, not a claim that this amount of data is sufficient. We also do not use the current results to claim generalization to arbitrary scenes or combinations of mechanisms.

  2. The distinction from effect-aware video editing is the input interface: those methods use an existing video, masks, or localization information, whereas we study erasure during generation from text prompts alone. We do not mean that those editing methods cannot remove physical effects.

Your questions about target-video screening and the training pipeline in Figure 2 also highlight details that need clearer explanation. Thank you for reading the paper carefully.

Reviewer_K1yqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper fine-tunes a text-to-video model to remove an object and the effects it causes. For example, a prompt describing a ball falling into water should produce water without the ball or splash. Another prompt showing the ball without interacting with the water should keep the ball visible. The proposed method, SRCD, changes the source object in training prompts while keeping the same target video, which shows neither the source object nor its effects. The model is trained to generate scenes without the source or its effects, even when the prompt requests them. The authors also introduce CES, a task-specific score for source and effect removal. The method is tested on seven types of physical interaction, with a separate adapter for each. SRCD slightly improves average CES over Matched, which uses the same training setup without replacing object names. However, the overall gain is not statistically significant after correction and human evaluation does not confirm it.

Critical strengths
  1. The paper addresses an interesting problem with three clear requirements: remove the source object, remove its physical effects and preserve its noncausal uses. For example, suppressing an object falling into water should also remove the splash and ripples, while keeping the same object when it appears without interacting with the water.
Critical weaknesses
  1. Table 1 reports the comparison between SRCD and external baselines. SRCD uses Wan 2.1 T2V 1.3B, while all external baselines use a different model, CogVideoX-2B. Although the authors acknowledge this limitation, these results cannot establish an advantage over existing methods.
  2. The authors introduce a novel metric for this task (CES). It is unclear how much a higher score reflects successful removal according to human judgments. Examples at different score levels and an analysis of the source/effect weighting would help clarify its meaning. Additionally, established video-quality metrics would help assess whether the method preserves generation quality.
Questions for authors
  1. Can the authors provide results in Table 1 with all methods on the same backbone with matched generation settings? Or alternatively, can the authors provide controlled evidence that any advantage of SRCD over the baselines persists across backbones?
  2. Can the authors provide a deeper analysis of the novel CES metric? Can the method be tested with established metrics for quality of generation?
  3. Can the authors report training time and hardware for their method? This would make comparison with other methods (particularly training-free ones) fairer.
  4. Does teacher distillation improve results over direct fine-tuning on the same target videos? A separate ablation would clarify its contribution.
1 comment
AuthorsOpenReview ↗

Response to Reviewer K1yq

Thank you for your review. We would like to clarify two points:

  1. The external baselines use a different backbone, so that comparison alone cannot establish which method is better. Our direct test of source replacement compares SRCD with Matched under the same backbone and training settings.

  2. CES measures source and effect removal; it does not replace an assessment of overall video quality. The paper also reports human evaluation and explicitly notes that it did not confirm the average gain seen in the automatic metric.

Your suggestions on the distillation ablation and training-cost reporting would help separate the contributions and costs of the individual components. Thank you for these concrete suggestions.

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

This paper studies causal-role erasure in text-to-video models: suppressing a source object and its dependent footprint in causal events while preserving the same noun in noncausal uses. It introduces Source-slot Randomized Counterfactual Distillation (SRCD), which trains mechanism-specific LoRA adapters on independently generated, screened source-free target videos while randomizing the source identity in the training prompt. Evaluation across seven physical mechanisms separates source removal, footprint removal, and noncausal preservation, and compares SRCD with matched training and external baselines. SRCD improves average graded erasure but does not establish overall superiority under corrected tests or human evaluation.

Critical strengths

The topic is interesting and timely.

Critical weaknesses
  • The experiments are clearly insufficient. Under the same backbone, SRCD does not significantly outperform Matched; the external baselines use a different backbone; and there is no per-case analysis of whether the same object (e.g., a knife) is wrongly erased in benign, noncausal contexts. Therefore, the claim that SRCD erases causal roles more precisely than existing methods is not well supported.
  • The motivation for causal-role erasure hinges on preserving noncausal uses; otherwise it is no different from standard concept erasure. However, the experiments do not really demonstrate this, and the qualitative results in Fig. 3 do not show preservation at all. From the presented results, the proposed method does not clearly outperform the baselines; T2VUnlearning also removes the source fairly cleanly.
  • The water-droplet example is not really relevant to harmful-content erasure. The motivation for erasure is to target harmful content, not to cherry-pick examples that look good, yet the paper keeps using this example throughout.
Questions for authors

see weakness

1 comment
AuthorsOpenReview ↗

Response to Reviewer KPCA

Thank you for reviewing our paper. We would like to clarify two points:

  1. Our focus is controllable generation, not only harmful-content erasure. The droplet and splash example illustrates that removing an object does not necessarily remove the effects it causes.

  2. Preserving the same object in noncausal uses is part of the task definition, and the paper includes paired-role tests. We do not take these results as evidence that SRCD outperforms existing methods; both the abstract and results report that the corrected tests and human evaluation do not confirm an overall advantage.

Your suggestion to include concrete noncausal examples would help illustrate this goal and the current results more clearly. Thank you for the suggestion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.