acceptodds
Under review as a conference paper at ICLR 2027

CleanVideo: Adaptive Subspace Intervention for Text-to-Video Concept Erasure

Abstract

Concept erasure in text-to-video diffusion models requires interventions that follow a concept as it emerges across spatial locations, video frames, and denoising steps. Fixed interventions can miss transient concept appearances or suppress unrelated content, compromising both erasure and temporal coherence. We introduce CleanVideo, a framework for selective concept erasure through adaptive interventions in low-dimensional hidden-state subspaces. A gating module jointly conditions on spatiotemporal visual features, denoising timesteps, and text semantics to control the location and strength of each intervention. We train the intervention modules with the pretrained diffusion backbone frozen, using target–surrogate prompt pairs to redirect unwanted concepts and preservation objectives to retain benign content. Experiments on CogVideoX-2B, CogVideoX-5B, and HunyuanVideo cover nudity, artistic style, and object erasure. Across these settings, CleanVideo improves the trade-off between concept suppression and generation quality relative to the evaluated baselines, while preserving visual fidelity and temporal coherence. Evaluations further measure video leakage whenever a target is detected in any sampled frame and assess concept recovery attacks with the intervention modules kept active. These results support adaptive representation intervention as an effective approach to selective video concept erasure.

Primary area:
generative models
Keywords:
concept erasure, text-to-video diffusion models, representation editing, subspace intervention, spatiotemporal adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MZTPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CleanVideo, a concept erasure method for text-to-video diffusion models. It uses low-dimensional subspace intervention with a tri-modal gating mechanism. The gating module combines visual features, timestep information, and text semantics to adjust the intervention strength. Meanwhile, the method uses surrogate concepts and preservation loss to reduce the influence on unrelated content. Experiments on nudity, artistic styles, and object erasure show that the method outperforms several baseline methods on three video generation models.

Critical strengths
  1. The proposed tri-modal gating mechanism considers spatial and temporal changes during video generation. It provides a reasonable way to adjust concept erasure at different regions and denoising steps.
  2. The experiments cover three erasure tasks and three video generation models. The results on multi-concept erasure and concept-recovery attacks are also useful.
Critical weaknesses
  1. Theorem A.4 uses a low-rank argument based on cross-attention, but it does not clearly explain why concept directions from different samples share the same low-dimensional subspace. Meanwhile, Eq. (3) uses a fixed orthonormal basis R, but its construction and relation to the target concept are not clearly explained.
  2. In Appendix E.11, removing a branch means replacing its output with a constant 0.5 after sigmoid. However, this also changes the scale of the final gate. The paper also lacks comparisons with simpler gating methods. Therefore, it is unclear whether the improvement comes from the three signals or the proposed fusion mechanism.
  3. The artistic style evaluation uses GPT-4o to classify the generated images into five artist styles. However, the evaluation prompt does not include a "none of the above" option. Even if the target style is successfully removed, the image must still be classified as one of these styles. This may affect the reliability of the style erasure results.
Questions for authors

Please see the weaknesses above. I am willing to reconsider my rating if these concerns are properly addressed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer MZTP

Thank you for the detailed feedback. We address your three concerns below.

W1. Construction of R and the cross-sample argument

R is initialized with orthogonal directions and kept fixed. Orthogonality avoids redundant directions in the limited update space. R determines where representations can change; target–surrogate pairs train A, b, and the gates to remove the target concept, while preservation examples constrain changes to benign content. Different samples share R, but receive different updates depending on their visual features, timestep, and text.

Sharing a subspace does not require identical concept directions across samples. In Appendix A, we now state the shared low-rank assumption explicitly: when the collected concept directions admit a low-rank approximation, their leading singular directions capture most of the total energy. The revised bound applies to the sample collection. This motivates low-dimensional intervention; the implemented R is fixed rather than constructed from these singular directions, and its use is learned through target–surrogate supervision.

W2. Gate strength and simpler alternatives

We added controls that replace each branch with a learnable constant vector. This lets the control learn its strength without using that branch’s input. We also tested fully constant gating and additive fusion. For additive fusion, all three branches are retained, with

g=σ((zvis+ztime+ztext)/3).g=\sigma((z_{\mathrm{vis}}+z_{\mathrm{time}}+z_{\mathrm{text}})/3).

All variants use the same training data and other training and evaluation settings on HunyuanVideo nudity erasure.

Default intervention strength: 1.0. Leakage is a percentage; quality scores use a 0–100 scale.

MethodLeakage ↓Object class ↑Subject consistency ↑Motion smoothness ↑Background consistency ↑
Full model7.596.296.995.797.8
Constant visual branch9.395.495.294.895.6
Constant temporal branch9.795.193.494.396.6
Constant separate text branch8.194.295.394.296.4
Additive fusion7.895.295.794.997.2
Fully constant gating11.694.894.694.596.8

We then varied inference strength without retraining:

MethodStrengthLeakage ↓Object class ↑Subject consistency ↑Motion smoothness ↑Background consistency ↑
Full model0.514.096.697.396.198.0
Full model1.07.596.296.995.797.8
Full model1.56.095.496.195.197.0
Additive fusion0.516.096.296.995.798.0
Additive fusion1.07.895.295.794.997.2
Additive fusion1.56.093.694.193.595.6
Fully constant gating0.518.096.096.695.998.2
Fully constant gating1.011.694.894.694.596.8
Fully constant gating1.58.092.893.093.195.0

At the same 6.0% leakage, the full model exceeds additive fusion by 1.4–2.0 points on all four preservation metrics. Thus, its advantage extends beyond a difference in intervention strength. These results are included in Appendix E.11, Tables 22–23.

W3. Forced-choice style evaluation

We added “none of the above” to the GPT-4o classifier and re-evaluated the same videos, keeping frame sampling and other settings unchanged. This option counts as no target-style detection on target prompts, but not as correct style preservation on non-target prompts.

Each cell below reports target-style detection ↓ / unrelated-style preservation ↑ (%), averaged over five artists.

MethodHunyuanVideoCogVideoX-2BCogVideoX-5B
Original model97.5 / 83.094.0 / 76.596.5 / 80.5
NegPrompt74.0 / 60.576.0 / 52.077.0 / 57.5
SAFREE55.0 / 52.058.5 / 45.057.0 / 50.5
VideoEraser30.5 / 70.038.0 / 61.538.5 / 67.0
T2VUnlearning25.0 / 74.534.5 / 66.032.0 / 71.5
CleanVideo16.5 / 80.528.0 / 74.024.5 / 78.0

CleanVideo retains the lowest target-style detection and highest unrelated-style preservation among the erasure methods on all three backbones. The results are included in Appendix E.7, Table 17; the original five-option results remain available for comparison.

AuthorsOpenReview ↗

Follow-up to Reviewer MZTP

Dear Reviewer MZTP, we would appreciate your feedback on our clarifications and new experiments. Do they address your concerns, or is there anything we can clarify further? Thank you.

My concerns have been largely addressed. I threfore decide to raise to raise my score.

AuthorsOpenReview ↗

Thanks to Reviewer MZTP

Thanks for your raising and reply! ! !

Reviewer_YZ8POpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

CleanVideo erases unwanted concepts like nudity, art styles, or specific objects from text-to-video models. It adds small low-rank adapters to a frozen model that steer the target concept toward a safe substitute, such as “clothed” instead of “naked.” A three-part gate controls them, using visual features, the denoising timestep, and the prompt, so editing happens only where and when the concept actually appears.

Critical strengths

Critical Strengths:

  1. Strong reported nudity-erasure results against the included baselines.

  2. Paper’s motivation for adapting the intervention according to the current visual representation, timestep, and prompt is technically sensible.

  3. The intervention gate design has an institutive interpretation, the textual branch estimates whether intervention is relevant, the visual branch estimates where target-related information appears, the temporal branch adjusts intervention according to the denoising stage.

Critical weaknesses

Critical Weaknesses:

  1. The paper lacks an overview figure of the pipeline, which makes Sec. 4 hard to follow. A diagram should show where the intervention modules sit in the transformer, how the three gating branches combine into g, and how training data, losses, and inference connect.

  2. Ring-A-Bell results are reported as a single number per backbone, without specifying which configuration (K16, K38, K77) was used or how prompts were sampled. Since robustness to adversarial prompts is a central claim, please report results separately for each Ring-A-Bell setting. Also, inference times and SSIM or LPIPS scores against the original video from the same seed should be reported.

  3. The paper uses GPT-5.1 to augment target prompts through lexical variation, scene diversification, and indirect expressions. It also constructs target-surrogate pairs and includes preservation prompts. These choices may substantially improve erasure performance independently of the gating mechanism. A useful ablation would compare: explicit target prompts only, target prompts with lexical augmentation, target prompts with scene diversification, target prompts with indirect expressions, the complete curation strategy. Without such comparisons, it is harder to determine how much of the improvement comes from the intervention architecture and how much comes from the training-data design.

Questions for authors

See the Weaknesses above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer YZ8P

Thank you for the feedback. We address each concern below.

W1. Pipeline overview

We added Figure 3 to show where the intervention sits in the frozen backbone, how the three gating branches combine, and how training data and losses connect to the intervention module. The diagram distinguishes forward computation from parameter updates.

W2. Ring-A-Bell settings, inference time, and perceptual preservation

Ring-A-Bell. We use K16, K38, and K77 with η=3, taking all 95 prompts from each configuration without additional sampling. Appendix E.2, Table 13 now reports results separately for these configurations. CleanVideo has the lowest leakage in all nine backbone–configuration combinations.

Inference time. Appendix E.3, Table 14 reports the SafeSora evaluation runtime: 47 minutes for CleanVideo versus 43 minutes for the original model, an increase of approximately 9.3%.

Perceptual preservation. We added LPIPS-AlexNet measurements between original and CleanVideo outputs on non-target prompts, using the same prompts, seeds, and generation settings. We average corresponding-frame distances within each video and report the mean ± standard deviation across video pairs.

BackboneVideo pairsFrames/videoLPIPS to original ↓
CogVideoX-2B176170.082 ± 0.031
CogVideoX-5B120170.088 ± 0.034
HunyuanVideo66170.095 ± 0.038

These results quantify changes relative to the original outputs and are included in Appendix E.1, Table 10.

W3. Contribution of prompt curation and gating

To examine what gating contributes with the data held fixed, we compared the full model with constant-branch controls, fully constant gating, and additive fusion. All variants use the same curated training data.

At default strength, the full model achieves lower leakage and higher preservation scores than all alternatives. At matched leakage, the comparison with additive fusion is:

MethodLeakage ↓Object class ↑Subject consistency ↑Motion smoothness ↑Background consistency ↑
Additive fusion6.093.694.193.595.6
Full multiplicative gating6.095.496.195.197.0

Both rows use HunyuanVideo nudity erasure at intervention strength 1.5. With the same data and the same leakage rate, the full model preserves content better across all four metrics. This shows an additional benefit from the gating design under the shared data pipeline. The comparisons are included in Appendix E.11, Tables 22–23.

AuthorsOpenReview ↗

Follow-up to Reviewer YZ8P

Dear Reviewer YZ8P, we have updated the paper with the additions described in our response. Please let us know if any concerns remain. Thank you for your time.

Reviewer_GUx7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Concepts in videos (T2V models) appear differently than images (T2I models). It appears in progression through frames and steps, and occupy different spatial regions in the frames. Existing method are not adaptive to these dynamic changes, it applies fixed rules- changes to text embeddings, or attention weights or uniform finetuning edits. The authors of CleanVideo first carefully curate the prompts to avoid under-erasure or over-erasure of concepts. Then, they form an equation to modify hidden representations inside the transformer blocks. The tri-modal gating feature is the one which is trainable (2-layer MLP with GeLU), and it covers spatial, temporal and text modalities. All the modality gates must reach substantial activations to obtain an activated final gate to avoid over-erasure of concepts in the video. For training, three losses- erasure, preservation and diversity loss is used. Both qualitative and quantitative results are compelling and comprehensive. CleanVideo effectively removes harmful concepts from all video frames without erasing unrelated concepts.

Critical strengths
  1. The authors correctly identify the need of adaptive erasure instead of the existing fixed erasure in video generation models due to the changing spatial, temporal features in videos, and therefore combines low-rank subspace editing with adaptive, trainable gates.

  2. The video generation backbone is fixed, only the intervention (gate) parameters change according to the training.

  3. CleanVideo clearly achieves the best results against all SOTA methods, with minimum video leakage.

  4. The experiments are exhaustive and ablations are comprehensive, across multiple metrics, models and benchmarks.

Critical weaknesses
  1. Ablations with respect to the number of prompts required for under/over-erasure was not provided. It would be useful to see how much the prompt capacity affects the erasure-preservation trade-off.
  2. The ablation on the ranks for nudity erasure was provided in the appendix. Was similar trend of used rank seen across other concept categories?
  3. Appendix E.3 reports the inference time, which was comparable to the other methods. However, the total training time and compute cost was not mentioned.
  4. In the whole paper, the effect of method on targeted concepts were illustrated through qualitative examples. It would be useful to see the effect on retain concepts as well.

Minor:

  1. The paper ICE[a] also shows representation/subspace-level erasure with applications in video unlearning. These baselines should also be included in evaluation.

References: [a] Das Biswas, Shristi, Arani Roy, and Kaushik Roy. "Now You See It, Now You Don't: Instant Concept Erasure for Safe Text-to-Image and Video Generation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer GUx7

Thank you for the helpful suggestions. We have added the following experiments and details.

1. Training data size

We tested 100, 250, and 500 target–surrogate pairs on HunyuanVideo, training each setting for 10 epochs.

Training pairsNudity leakage ↓Object class ↑Subject consistency ↑Motion smoothness ↑Background consistency ↑
10015.296.096.495.297.4
2509.496.196.795.597.6
5007.596.296.995.797.8

Under this fixed-epoch schedule, more training pairs improve erasure while preserving quality. Leakage drops by 5.8 points from 100 to 250 pairs and a further 1.9 points at 500 pairs. These results are included in Appendix E.14, Table 26.

2. Rank sensitivity beyond nudity

We added a rank ablation on CogVideoX-2B, averaged over five artist styles. Evaluation allows the classifier to select “none of the above.”

RankTarget-style detection ↓Unrelated-style preservation ↑Trainable gating parameters
148.578.20.8M
239.077.51.6M
428.074.03.2M
826.571.86.4M

Increasing rank from 1 to 4 substantially improves erasure. Increasing it to 8 gives only a further 1.5-point reduction in detection, while preservation drops by 2.2 points and gating parameters double. This supports rank 4 as a practical balance for this setting. See Appendix E.10, Table 20.

3. Training cost

We added estimated per-concept training costs in Appendix D.5, Table 7. Estimates use measured step times from short runs on one A100 80GB GPU, with batch size 1, 17 frames, and 320×320 resolution, extrapolated to 500 target–surrogate pairs over 10 epochs.

BackboneEstimated training time per concept
CogVideoX-2B2.5 h
CogVideoX-5B6.3 h
HunyuanVideo3.1 h

With one GPU, these values also correspond to GPU-hours. They cover training iterations and exclude model loading, data preparation, checkpoint saving, and evaluation.

4. Preservation examples

Figure 2 and Appendix Figures 17 and 21 show target removal alongside retained scene content. In particular, the chainsaw and golf-ball examples illustrate preservation of surrounding scene elements after removing the target object. These examples complement the quantitative preservation metrics.

5. Comparison with ICE

We reproduced ICE from its paper and supplementary material and evaluated CogVideoX-2B nudity erasure using the same test prompts, generation settings, and evaluation protocol as CleanVideo.

MethodSafeSora leakage ↓Ring-A-Bell leakage ↓Object Class ↑
Original model20.3%32.7%88.8
ICE (our reproduction)8.6%9.7%85.2
CleanVideo1.7%3.5%87.9

CleanVideo lowers leakage by 6.9 and 6.2 percentage points and improves Object Class by 2.7 points. ICE offers the advantage of training-free weight editing. The comparison is included in Appendix E.15, Table 27.

Thank you for the clarifications. I will keep my score.

AuthorsOpenReview ↗

Thanks to Reviewer GUx7!

Thank you for your feedback and for supporting our work and rebuttal!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Discussion

We thank the reviewers for your helpful suggestions. During this rebuttal period, we:

  1. Clarified the construction of R and revised the cross-sample theoretical analysis.
  2. Added a pipeline diagram showing the intervention module, gating branches, and training supervision.
  3. Added learnable-constant gating controls and an additive-fusion baseline, followed by an intervention-strength sweep.
  4. Added a “none of the above” option to the style evaluator and re-evaluated all methods.
  5. Added training-data-size and artistic-style rank ablations.
  6. Reported per-concept training-cost estimates.
  7. Reproduced ICE and added a comparison on CogVideoX-2B.
  8. Added Ring-A-Bell results by configuration and same-seed LPIPS measurements.

All updates are included in the revised paper. We received positive feedback from two reviewers: MZTP responded positively to our rebuttal and raised their score, while GUx7 reaffirmed their support for acceptance after reviewing our clarifications. Our comparisons with training data held fixed also provide evidence of the gating design’s additional contribution, addressing YZ8P’s concern about the source of the gains.

We thank all reviewers, ACs, SACs and PCs again for your time and helpful feedback!!!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.