acceptodds
Under review as a conference paper at ICLR 2027

EPO: Concept Erasing in Text-to-Video Models via Energy-based Erasing Preference Optimization

Abstract

Modern Text-to-Video (T2V) models generate increasingly realistic and coherent videos, but they also risk producing undesired content such as nudity, personal identities, or unsafe objects. Existing concept erasure methods, largely developed for Text-to-Image models, are difficult to transfer directly to T2V generation because undesired concepts can reappear through diverse appearances, motions, viewpoints, and temporal compositions. In this paper, we propose EPO, an energy-based preference optimization framework for T2V concept erasure. Instead of relying on local suppression of limited samples, EPO formulates erasure as energy-based preference shaping over the open-ended video generation space. We introduce an Energy-based Erasing Preference and instantiate it with an energy discrepancy objective that enlarges the energy gap between desired and undesired generations. To mitigate catastrophic forgetting caused by broad energy suppression, we further introduce a Manifold-guided Target Anchoring strategy, which redirects undesired prompts toward nearby safe semantic anchors while preserving spatiotemporal generation priors. Extensive experiments demonstrate that EPO effectively suppresses undesired concepts and better preserves video generation quality. Anonymous code is available at https://anonymous.4open.science/r/E-2PO-AnonymousRelease.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Concept Erasing, Text-to-Video Generation, Preference Optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RSrrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes E²PO, an energy-based preference optimization framework for concept erasure in text-to-video (T2V) models. The method introduces a multi-negative energy discrepancy objective to suppress undesirable generations and a Manifold-guided Target Anchoring (MTA) strategy to preserve unrelated spatial and temporal semantics. Experiments on nudity, identity, and object erasure demonstrate improved erasure effectiveness, adversarial robustness, and video generation quality compared with existing baselines.

Critical strengths
  1. The problem is essential. The paper addresses the challenges of concept erasure in T2V models, where undesirable concepts are entangled with motion, temporal dynamics, and scene structure.

  2. The method is well-motivated. Combining multi-negative preference optimization with semantic target anchoring provides a reasonable approach to balancing erasure effectiveness and preservation of generation capabilities.

  3. The experimental results are convincing. E²PO consistently outperforms representative baselines across multiple erasure tasks. The extensive ablations, robustness evaluations, and additional analyses in the appendix further support the effectiveness of the proposed method.

Critical weaknesses
  1. The clarity of methodological novelty is limited. Although E²PO is motivated by energy-based distribution shaping, its practical objective closely resembles existing multi-negative contrastive preference optimization. The unique contribution of the energy-based formulation remains insufficiently distinguished from conventional preference optimization.

  2. The ablation study shows that the null target achieves stronger erasure effectiveness at the cost of slightly lower generation quality. Does this suggest that MTA primarily shifts the erasure–preservation trade-off rather than providing a substantive improvement? This raises questions about the necessity and unique contribution of MTA.

Questions for authors
  1. Could the authors further clarify the essential methodological distinction between E²PO and conventional multi-negative contrastive preference optimization, particularly the unique benefits introduced by the energy-based formulation?

  2. Could the authors further justify why semantic proximity in the text-conditioning space is a reliable criterion for preserving video motion and temporal structure?

  3. Following 2, could the authors clarify whether MTA provides benefits beyond adjusting the erasure–preservation trade-off, particularly given the stronger erasure effectiveness achieved by the null target?

  4. (Minor) About comparison with recent baselines. The evaluated baselines are primarily from 2024–2025. Could the authors briefly discuss the availability of more recent T2V concept erasure methods and clarify the rationale behind the selected baselines?

I am open to raising my score if the authors can adequately address these concerns during the rebuttal.

Reviewer_Yj5aOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies concept erasure for text-to-video models. The authors argue that erasure methods designed for text-to-image models do not work well on video, because an unwanted concept in a video can come back through many different looks, viewpoints, motions, and time orders. They propose E2PO, which treats erasure as preference optimization with a KL term and writes the target preference in an energy-based form. In practice, the model is trained with an offline loss that compares one desired video with several undesired videos for each erasing prompt, and the reward is estimated from velocity errors along the flow path, following Diffusion-DPO.

The desired video is made by editing a normal generation toward a safe anchor concept with a training-free flow editing method, so that the layout and motion stay the same. The anchor is chosen by a strategy called Manifold-guided Target Anchoring (MTA), which finds the closest safe concept in the T5 text feature space using averaged prompt features. Experiments on Wan2.1-1.3B and CogVideoX-2B cover nudity, identity, and object erasure, plus Ring-A-Bell and MMA-Diffusion attacks and a user study. E2PO gets the lowest nudity rate and the best balance between erasure and quality among the T2V baselines, and the ablation studies look at the anchor choice, the way negative samples are chosen, the number of negatives M, and β.

Critical strengths
  1. The paper studies an important problem. Concept erasure for text-to-video models has received much less attention than for text-to-image models, and safety risks increase as video models become stronger and more widely used.

  2. The method is simple and easy to use. It is based on a well-known preference optimization loss, only needs offline training data generated by the model itself, and is tested on two different T2V models.

  3. The results are strong compared with the existing T2V erasure baselines. E2PO has the lowest nudity rate on both models (3.34% on Wan and 10.75% on CogVideoX) while keeping VBench scores close to the original model, and it also gets the best erasure and preservation scores on the identity and object tasks.

  4. Building the desired video by editing a normal generation toward a safe anchor is a reasonable idea. It gives the model a clear target that keeps the layout, motion, and scene of the original video, which helps reduce damage to other content during erasure.

Critical weaknesses
  1. I may be missing a detail, but the connection between the theory and the practical loss is not entirely clear to me. Theorem 4.2 assumes positives from an ideal erasing distribution and a suitable perturbation kernel. Could the authors clarify how the edited positives and generated harmful negatives approximate this setup?

  2. I would have liked more direct validation of the video-specific motivation. The evaluated concepts seem mainly appearance-based, and the metrics largely rely on frame-level recognition. Action concepts or video-level leakage and temporal recurrence metrics could provide useful additional evidence.

  3. The appendix provides helpful Flow-DPO and anchor comparisons. A supervised fine-tuning baseline using the same edited positives might further clarify the benefit of preference optimization. Comparing MTA and null targets at similar erasure strength could also help assess the preservation advantage.

  4. A few evaluation details could be clearer. Additional motion measures, confidence intervals, and independent training runs would help assess the results. I may have overlooked these details, but could the authors clarify the adaptive attack settings and which unaffected classes are evaluated in the ten-concept erasure setting in Table 23?

Questions for authors

Please refer to the Weaknesses. I am open to adjusting my score based on the authors’ response.

Reviewer_hE3DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes E2PO, a text-to-video erasing method based on energy preference. It supresses unwanted concepts by using the energy preference of multiple negative samples. Meanwhile, this paper proposes MTA (Manifold-guided Target Anchoring), which selects safe concepts with similar semantics to construct positive samples, avoiding the influence of quality of unrelated concepts. Experiments on erasing nudity, identities, and objects show that the method outperforms multiple baseline methods.

Critical strengths
  1. The combination of multi-negative preference optimization and MTA provides a reasonable way to suppress unwanted concepts while preserving unrelated content.
  2. The experiments cover three erasure tasks and multiple video generation models. The comparison with multi-negative Flow-DPO in Appendix D.4 is also useful.
Critical weaknesses
  1. The objective in Eq. (6) is related to existing multi-negative contrastive and preference optimization losses. Although Theorem 4.2 provides a result, it is unclear whether the practical implementation satisfies the assumptions. Therefore, the real contribution of the theorem remains unclear if it is not aligned with implementation.
  2. The MTA strategy selects safe anchors based on distances in the frozen text encoder space. However, similar semantics are not necessarily equal to similar video structure or motion. For example, Table 16 reports a RAFT flow cosine of only 0.354 for nudity pseudo-positive samples. Therefore, it remains unclear how reliably the selected anchors preserve the original spatiotemporal information.
  3. The computional cost of constructing preference data is not reported.
Questions for authors

Please see the weaknesses above. I an willing to keep positive rating if the weaknesses and concers are properly addressed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.