acceptodds
Under review as a conference paper at ICLR 2027

FOMO: Forget the Concept, Don't Miss Out on the Scene in Selective Video Unlearning

Abstract

The rapid advancement of generative video models has enabled the synthesis of increasingly realistic and temporally coherent videos, while also raising concerns about the generation of harmful content. The reliance on large-scale web datasets during training inevitably exposes these models to undesirable material, making concept unlearning an essential mitigation. Existing methods mainly target static visual concepts, such as objects, identities, or unsafe appearance, largely overlooking motion unlearning. Furthermore, these approaches often pay little attention to preserving the surrounding scene. As a result, successful concept removal may unintentionally alter the background, composition, or overall video dynamics. We argue that effective unlearning should ideally change only what is targeted, while minimizing unnecessary changes to the remaining scene. In this work, we introduce FOMO, to the best of our knowledge the first training-based selective video unlearning method that directly treats preservation of the original scene as a priority. We formulate unlearning around two complementary objectives: what to change and what to preserve. Our method localizes concept-related representations and modifies them, while the preservation mechanism maintains non-target scene information without requiring auxiliary data. Beyond simply erasing unwanted concepts, FOMO explicitly redirects the generation toward a specified safe alternative. We further extend this formulation to motion unlearning, where the concept is defined by temporal behavior rather than a fixed spatial region. Our solution achieves effective unlearning across unsafe content, object, and motion concepts, while achieving the best trade-off between concept removal and scene preservation. [Anonymous page with videos](https://anonymous-73915942.github.io/fomo-video-unlearning/)

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
video generation, machine unlearning

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

FOMO aims to erase target concepts from text-to-video models while preserving the original scene. It constructs spatiotemporal masks using attention anchors and representation similarity, then trains LoRA adapters through source–safe representation alignment and masked value preservation. CHI-based head selection extends the method to motion unlearning. Experiments on HunyuanVideo cover unsafe content, objects, public figures, and motion.

Critical strengths
  1. Scene preservation is a practically important problem in video concept unlearning. I think the focus on reducing unnecessary scene changes, together with the extension to motion, is well motivated;
  2. The method makes a clear connection between localization, representation editing, and preservation. Training through LoRA without auxiliary retain data is also a useful practical feature;
  3. The experiments cover several tasks and datasets. I particularly appreciate the comparison between ( LHL_H ) and ESD at the same Unsafe Rate in Table 6: the better scene similarity provides direct support for representation-level intervention.
Critical weaknesses
  1. I think the claim of the best erasure–preservation trade-off needs stronger support. DINO and LPIPS are computed over full frames, so residual target content can also improve these scores. This matters in Table 1, where FOMO and ESD have Unsafe Rates of 19.54% and 1.76%, respectively. The equal-rate comparison in Table 6 is helpful, but covers one ablation setting on GEN. Comparing non-target content at similar erasure rates would make the claim more convincing;
  2. The motion results show substantial target suppression. My main concern is whether other actions remain intact. Tables 5 and 15 compare only with the base model, while the broader VBench preservation test concerns nudity. For running to walking, checking whether the same adapter generates walking and other retained actions would be more informative about selectivity than motion smoothness. Including video unlearning baselines and evaluating both safe motions and retained actions would better establish the contribution of the motion branch;
  3. Eq. 2 modifies the full attention representation within the mask, while Eq. 4 mainly constrains representations outside it. Yet nudity or motion can share video tokens with identity, appearance, or other behaviors that should be retained. The method therefore appears to rely on the source–safe feature difference mainly capturing the target attribute change. I would appreciate a more explicit justification of this assumption, particularly given the dependence on safe mappings in Appendix G;
  4. CHI measures differences across frame representations, which can also arise from camera motion or background changes. I am not sure this is sufficient to identify heads specific to the target action. There is also a distinction between selecting heads for ( LHL_H ) and restricting updates to those heads. Appendix A places LoRA in text Q/K projections without a corresponding head restriction. Clarifying what head selection actually constrains would help explain the scope of the motion-selective intervention.
Questions for authors
  1. Does FOMO retain its preservation advantage when non-target content is evaluated at comparable erasure rates?
  2. How well does the same adapter generate the safe motion and other retained actions after one action is erased?
  3. How does the source–safe feature difference limit the edit to the target attribute while preserving other attributes within the mask?
  4. Does head selection restrict the loss computation or the heads affected by the updates? How does CHI distinguish the target action from other temporal changes?
4 comments

W1/Q1: We thank the Reviewer for this suggestion. We agree that full-frame DINO and LPIPS may favor methods that retain target content, and that preservation should be compared at similar erasure rates.

To address this, following HATIE [2], we use segmentation to identify subject and background regions in videos from the GEN dataset and compute DINO, CLIP, LPIPS, and L2 separately for each region, denoted as s- and b-metrics, respectively. We also use Qwen3-VL-8B-Instruct to evaluate subject identity and appearance preservation, as well as unrelated background preservation, following [1]. Specifically, we ask two questions:

  1. Is the subject’s identity and overall appearance preserved, allowing changes to clothing and skin exposure needed to remove nudity?

  2. Is unrelated background content preserved, allowing local changes in pose or interaction needed to remove nudity?

We report Both Preserved when both components are preserved and Joint Success when both are preserved and the generated video is classified as safe. We additionally evaluate FOMO with different preservation weights (λV=0.2,0.5,0.7\lambda_V = 0.2, 0.5, 0.7) and ESD at checkpoints 10, 30, and 50 to compare methods at similar Unsafe Rates (see Table C.1 below). As shown in Fig. 14 in paper, at 26-30% Unsafe Rate, FOMO (λV=0.5\lambda_V =0.5) preserves both the subject and background in 90% of videos, compared with 32% for ESD and 26% for T2VUnlearning. FOMO also achieves lower b-LPIPS (0.0778 vs. 0.1248 and 0.1993, respectively).

At 4% Unsafe Rate, the alignment-only variant (LHL_H) also outperforms ESD in preservation (50% vs. 14% Both Preserved; 0.0942 vs. 0.2210 b-LPIPS). These results suggest that FOMO can better preserve non-target content at comparable erasure rates.

Table C.1: Erasure-preservation trade-off on the GEN dataset.

Method / VariantUnsafe Rate (%) ↓b-DINO ↑b-CLIP ↑b-LPIPS ↓b-MSE ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
Negative Prompt820.85180.93500.18410.01833236306
SAFREE400.83920.93750.18800.022228242214
T2VUnlearning300.86070.94510.19930.026626302614
ESD (10)300.83920.95250.12480.009936403230
ESD (30)160.80010.92430.18750.016826302222
ESD (50)40.78100.91430.22100.021814181414
LHL_H40.89860.96260.09420.007762665048
LVmaskL_V^{\mathrm{mask}}80.001.00001.00000.00000.000010010010020
LH+LVfullL_H + L_V^{\mathrm{full}}440.91920.97130.07680.005694929046
FOMO (λV=0.2\lambda_V=0.2)200.90980.96190.08420.007480868060
FOMO (λV=0.5\lambda_V=0.5)260.91760.96910.07780.005990929064
FOMO (λV=0.7\lambda_V=0.7)380.91900.97220.07220.005692929254

[1] Wang et al., “Rethinking Real Image Editing: Unleashing Diverse Editing Operators via Multi-Objective Optimization,” WACV 2026.

[2] Ryu et al., “Towards Scalable Human-aligned Benchmark for Text-guided Image Editing,” CVPR 2025.


W2/Q2: We conducted new experiments to test whether each unlearning adapter could still generate the other five actions. We used the same protocol as for Motion Rate, including the same action-presence question and Qwen3-VL-32B-Instruct evaluator. The results are shown below and in Table 22 in the paper. Across the 30 action pairs, FOMO generated the requested action more often than ESD (84.67% vs. 67.00%), while the base model achieved 96.67%.

Table C.2: Preservation of non-target motions after unlearning.

Erased MotionMethodRunningFightingTongueDancingJumpingKnifeAvg. ↑
RunningBase×951001008510096
ESD×2010040804056
FOMO×80100100808088
FightingBase100×1001008510097
ESD100×10020906074
FOMO100×10060808084
Sticking out the tongueBase10095×1008510096
ESD10060×1001006084
FOMO10080×1001009094
DancingBase10095100×8510096
ESD10030100×404062
FOMO10040100×608076
JumpingBase10095100100×10099
ESD102010040×8050
FOMO405010080×8070
Brandishing a knifeBase1009510010085×96
ESD1002010060100×76
FOMO10080100100100×96
Overall Avg.Base——————96.67
ESD——————67.00
FOMO——————84.67

W3/Q3: We thank the Reviewer for raising this point. We agree that the target concept can share tokens with identity or appearance, but FOMO limits unintended changes in three ways:

  1. Minimal prompt change: We change only the target phrase from cc to c’c’, for example nude → clothed. We use the same latent state and frozen model for both prompts, so the difference between their features primarily reflects the change to the target phrase.

  2. Localized unlearning: We compute soft mask MM in Eq. (2) which controls the edit strength at each token. For motion, we use only the five highest-CHI heads per block to build mask and compute LHL_H. This improves joint subject and background preservation from 51.67% to 59.17% compared with using all heads (see Table 8 in paper).

  3. Scene preservation: In Eq. (4), LVL_V encourages the model to retain features from the original prompt cc at the first denoising step, with weight (1−M)(1-M). Tokens with (0<M<1)(0<M<1) receive both the removal and preservation signals. For example, if removing the target concept slightly changes a person’s face, LVL_V encourages those facial features to stay close to their original appearance.

For nudity unlearning, depicting the person clothed is a natural safe alternative [1]. We replace nude with clothed and keep the rest of the prompt fixed. Table 23 shows similar Unsafe Rates across subjects and settings for this contrast.

For objects, the replacement can affect the whole scene. LHL_H guides the edit toward the replacement in c′c', while LVL_V preserves the surroundings and thereby favors replacements that fit the scene. We choose a safe concept that can play a similar role in the same setting. For example, a rubber duck can sit on the same grassy field as a golf ball, so both losses support the edit.

When c′c' is empty, LVL_V still anchors the scene to the original prompt cc, while the model has more freedom inside the mask. Fig. 15 shows both outcomes. The golf ball disappears and the grassy field remains, whereas replacing the parachute with a boat shifts the scene to water. These examples suggest that object unlearning better balances target removal and scene preservation when the safe concept plays a similar role in the original setting.

[1] Facchiano et al. Video Unlearning via Low-Rank Refusal Vector. ICLR 2026.


W4/Q4: CHI measures how strongly each attention head responds to changes across frames. Since camera movement also introduces changes between frames, it may influence CHI scores. During training, we prompt the model to generate the target action without explicitly requesting camera movement. We also clarified in Appendix A that LoRA adapters are trained on the same Q/K projections for all concepts, while CHI selects the attention head outputs used to compute LHL_H.

Additionally, we evaluated several head configurations for motion unlearning (see Table C.3. below). Following W1/Q1, we asked Qwen two questions:

  1. Is the subject’s identity and overall appearance preserved, allowing for pose or interaction changes needed to remove the action?
  2. Is unrelated background content preserved, allowing for pose or interaction changes needed to remove the action?

CHI top-5 achieves the lowest Target Motion Rate (12.50%) and the highest Joint Success (50.83%), which requires removing the target action while preserving both the subject and the background. It also achieves the best background preservation, suggesting that removing the target action has limited impact on camera motion.

Table C.3: Effect of attention head selection on motion unlearning and scene preservation.

MethodMotion Rate (%) ↓DINO ↑LPIPS ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
Base96.67——————
ESD13.330.42970.57146.6718.336.675.00
FOMO (top-5)12.500.70130.346761.6773.3359.1750.83
FOMO (all heads)15.830.65370.379853.3366.6751.6744.17
FOMO (random-5)25.830.60910.392247.5064.1745.8338.33
FOMO (top-10)16.670.63030.395050.0062.5047.5042.50

Thanks for the reply and for running the extra experiments. I was especially interested in whether the same adapter could still generate other actions after unlearning one, so I’m glad you tested that directly. The comparison at similar erasure rates and the explanation of head selection were helpful too.

I’ll keep my positive score. Good luck with the paper!

Dear Reviewer iomB,

We are pleased that we have addressed most of your concerns. Thanks for your insightful comments and approval of our work!

Reviewer_E2eqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FOMO, a training-based approach to selective concept unlearning in text-to-video Diffusion Transformers. FOMO uses a frame-wise video-text attention anchor expanded by similarity among attention-output representations to localize a target concept. A frozen teacher produces source and safe-prompt representations with a LoRA-trained student aligned to their masked interpolation. A complementary, non-target masked value-preservation loss, applied at a high-noise timestep, seeks to retain the rest of the scene without an auxiliary retain set. For motion concept selection, the proposed method restricts alignment to attention heads selected by the temporal Calinski-Harabasz index. Experiments evaluate the proposed method on HunyuanVideo across unsafe-content, object, public-figure, and motion unlearning.

Critical strengths
  1. Meaningful problem and a well-motivated erasure-preservation objective: Selective video unlearning is a challenging task because changing a concept may disrupt the surrounding spatial and temporal scene. The pairing of localized teacher-student alignment with masked value preservation explicitly addresses this problem. Table 6 is informative: on GEN, erasure-only L_H yields 4% Unsafe Rate with DINO 0.5412, while masked preservation yields 26% with DINO 0.6832; global preservation increases Unsafe Rate to 44%. This supports the motivation for excluding the target region from preservation.

  2. Broad coverage of unlearning tasks: The evaluation cover unsafe-content, object, public-figure, and motion unlearning using HunyuanVideo backbone. For example, Table 3 reports 92.41% ESR-1 for object erasure while retaining 66.62% PSR-1, and Table 15 provides per-motion outcomes and quality scores. It makes the proposed problem setting relevant beyond a single unsafe-content benchmark.

Critical weaknesses
  1. The central claim of preserving non-target scene content is not directly measured. DINO and LPIPS are computed between whole corresponding frames, and VBench assesses general video properties. None of them specifically verifies unchanged non-target regions, object identities, spatial relations, or temporally tracked backgrounds after the target is edited. The qualitative examples are suggestive but cannot replace a region-aware or human-verified preservation study. Moreover, FOMO's 19.54% nudity Unsafe Rate is much higher than ESD's 1.76% (Table 1). The ablation results indicate that erasure-only alignment can reach 4% on GEN (Table 6). The asserted best trade-off therefore needs stronger Pareto-style evidence at comparable erasure strengths, not only isolated operating points.

  2. The evidence for the motion-specific contribution is insufficient. The paper adopts CHI-based temporal-head ranking from prior interpretability work (Sec. 3.4), but provides no quantitative all-head vs. CHI-selected vs. random-head ablation, sensitivity to the number of selected heads, or motion-task comparisons with adapted existing erasure methods. As a result, the reduction in Target Motion Rate (Table 5) cannot be attributed specifically to selecting motion-sensitive heads rather than the generic source-to-safe representation alignment. Clarifying the consistency of the evaluated motion list (Sec. 4 mentions 'walking'; Appendix B.4/Table 15 instead includes 'brandishing a knife') would also help interpret the results.

  3. Experiments are based on only the HunyuanVideo backbone. This is a substantial limitation for a method whose claims concern selective video unlearning more generally. It is not clear whether the proposed method can be transferred to other video diffusion-transformer architectures or models with different text–video attention structures.

Questions for authors
  1. Can the authors provide an erasure-versus-preservation curve (e.g., varying alpha and lambda_V) and non-target-region preservation measurements, ideally using tracked regions, masks, and/or human judgments? How does FOMO compare with ESD and L_H-only under matched unsafe/target-motion rates rather than at one operating point?

  2. On the same six motion prompts, how do (i) all heads, (ii) randomly chosen heads, and (iii) CHI-selected top-k heads compare, including different k values? Can the authors provide an appropriate motion-unlearning baseline and state the evaluation sample count / uncertainty for each motion?

  3. Does the proposed method relies on specific architectural details, including double-stream and single-stream blocks, attention projections, and selected layer indices? Is the proposed method architecture-independent? How about evaluating the proposed method with different backbones?

5 comments

We thank the Reviewer for highlighting the importance of preserving non-target content and for recognizing the breadth of our video unlearning evaluation.

W1/Q1: We agree that evaluating subject and background regions separately allows a more detailed assessment of scene preservation.Following HATIE [2], we segment these regions in GEN videos and compute DINO, CLIP, LPIPS, and MSE separately for each region, denoted by the prefixes s- and b-, respectively. Inspired by the binary VLM evaluation [1], we use Qwen3-VL-8B-Instruct to compare each edited video with its reference and answer two questions:

  1. Is the subject’s identity and overall appearance preserved, allowing changes to clothing and skin exposure needed to remove nudity?

  2. Is unrelated background content preserved, allowing local changes in pose or interaction needed to remove nudity?

Both Preserved denotes the percentage of videos that pass both checks. Following the joint scoring approach in [1], we also report Joint Success, the percentage of all videos that are classified as safe and pass both preservation checks.

We evaluated two additional FOMO variants (λV=0.2\lambda_V=0.2 and 0.70.7), alongside the original (λV=0.5\lambda_V=0.5) setting. Fig. 14 illustrates the relationship between Unsafe Rate and both VLM-based preservation (Both Preserved) and b-LPIPS.

On GEN, FOMO (λV=0.5\lambda_V=0.5) achieves a 26% Unsafe Rate, comparable to 30% for both ESD (10) and T2VUnlearning (see Table B.1 below). At these rates, FOMO preserves both the subject and background in 90% of videos, compared with 32% for ESD (10) and 26% for T2VUnlearning. FOMO also achieves a lower b-LPIPS (0.0778 vs 0.1248 and 0.1993, respectively).

At the same 4% Unsafe Rate, the alignment-only variant (LHL_H) preserves both the subject and background more than ESD (50) (50% vs. 14%), with lower b-LPIPS (0.0942 vs. 0.2210).

Table B.1: Erasure-preservation trade-off on the GEN dataset.

Method / VariantUnsafe Rate (%) ↓b-DINO ↑b-CLIP ↑b-LPIPS ↓b-MSE ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
Negative Prompt820.85180.93500.18410.01833236306
SAFREE400.83920.93750.18800.022228242214
T2VUnlearning300.86070.94510.19930.026626302614
ESD (10)300.83920.95250.12480.009936403230
ESD (30)160.80010.92430.18750.016826302222
ESD (50)40.78100.91430.22100.021814181414
LHL_H40.89860.96260.09420.007762665048
LVmaskL_V^{\mathrm{mask}}80.001.00001.00000.00000.000010010010020
LH+LVfullL_H + L_V^{\mathrm{full}}440.91920.97130.07680.005694929046
FOMO (λV=0.2\lambda_V=0.2)200.90980.96190.08420.007480868060
FOMO (λV=0.5\lambda_V=0.5)260.91760.96910.07780.005990929064
FOMO (λV=0.7\lambda_V=0.7)380.91900.97220.07220.005692929254

[1] Wang et al., “Rethinking Real Image Editing: Unleashing Diverse Editing Operators via Multi-Objective Optimization,” WACV 2026.

[2] Ryu et al., “Towards Scalable Human-aligned Benchmark for Text-guided Image Editing,” CVPR 2025.


W2/Q2: We conducted ablations across six motions (Table B.2), comparing four attention head selection variants: top-5, top-10, all heads, and random-5. We followed the Qwen evaluation protocol described in W1/Q1, using two prompts adapted to motion removal:

  1. Is the subject’s identity and overall appearance preserved, allowing for pose or interaction changes needed to remove the action?

  2. Is unrelated background content preserved, allowing for pose or interaction changes needed to remove the action?

The top-5 variant achieves the lowest Motion Rate (12.50%), indicating the most effective removal of the target motion. It also best preserves the subject's appearance (Subject Appearance: 61.67%) and background (Background Appearance: 73.33%). Using more attention heads leads to greater changes in scene content. In contrast, the lower performance of random-5 highlights the importance of CHI-based head selection.

We evaluated 120 videos per method (20 per motion), using identical prompts and random seeds, including for ESD. We added the sample count to Appendix B.4.

In Section 4, walking was mistakenly listed instead of brandishing a knife. We thank the Reviewer for identifying this inconsistency, which we corrected in the main paper to match Appendix B.4 and Table 18.

Table B.2: Effect of attention head selection on motion unlearning and scene preservation.

MethodMotion Rate (%) ↓DINO ↑LPIPS ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
Base96.67——————
ESD13.330.42970.57146.6718.336.675.00
FOMO (top-5)12.500.70130.346761.6773.3359.1750.83
FOMO (all heads)15.830.65370.379853.3366.6751.6744.17
FOMO (random-5)25.830.60910.392247.5064.1745.8338.33
FOMO (top-10)16.670.63030.395050.0062.5047.5042.50

W3/Q3: FOMO relies on attention mechanisms rather than a specific transformer block structure. Concept localization uses interactions between video queries and text keys, while the alignment loss operates on per-head attention outputs before the output projection.

To evaluate transferability, we additionally applied FOMO to CogVideoX-2B, which uses 30 joint-attention blocks, unlike HunyuanVideo's 20 double-stream and 40 single-stream blocks. While HunyuanVideo applies alignment and preservation losses to separate block types, CogVideoX-2B applies both within the same stack, using 17 IMAP-selected blocks for alignment and the last six for preservation. LoRA is applied to the text-stream query and key projections in HunyuanVideo and to the joint-attention query and key projections in CogVideoX-2B.

For object unlearning, Table 4 shows that FOMO improves ESR-1 from 55.09 to 87.68 and ESR-5 from 20.15 to 62.68, while maintaining preservation close to the original model (PSR-1: 45.32 vs. 44.91; PSR-5: 76.64 vs. 79.85). SAFREE achieves comparable erasure (ESR-1: 89.41) but substantially lower preservation (PSR-1: 29.13; PSR-5: 52.59).

These results support FOMO's transferability across the evaluated joint-attention architectures. We added the results to the main paper (Table 4), with additional experiments and implementation details in Appendices D and A.

Dear Reviewer E2eq,

Thank you again for your constructive feedback. We hope our rebuttal addresses your concerns. If you have any further questions, we would be happy to answer them.

Reviewer_CS7pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents FOMO, a method for suppressing selected concepts and actions in video generation while preserving other scene content. It trains LoRA adapters on HunyuanVideo by guiding attention features toward those produced by a prompt containing a safe replacement concept. Localization masks and a preservation loss help limit changes to unrelated content. For action removal, the method selects motion-sensitive attention heads and applies the removal objective to their outputs. The evaluation covers unsafe content, objects, public figures, and actions, with quantitative comparisons, ablations, and video examples.

Critical strengths

Removing an unwanted action while keeping the rest of the scene is a useful goal. The method directly addresses this problem, and the demo videos show that it can suppress the target action in several cases. Extending concept removal to actions makes the work interesting.

The experiments provide evidence that the method can reduce unwanted changes. In Table 6, the version using only the attention-feature loss achieves the same 4% unsafe rate as ESD, with substantially better DINO and LPIPS scores. This is a useful comparison because the preservation improvement is achieved without weakening target removal in that setting.

Critical weaknesses

It is unclear how often action removal preserves the subject and background. In some demo videos, the target action is suppressed, but the subject or background also changes substantially. Appendix H acknowledges that some contextual changes may be necessary. However, the average DINO and LPIPS scores do not tell us how often unrelated content changes unnecessarily. Since scene preservation is a central goal, I would like a clearer breakdown of successful and unsuccessful cases. Showing the localization masks for examples with large scene changes would also help check whether they include unrelated regions.

The experiments do not clearly show what the motion-specific design adds. Table 5 compares FOMO only with the original model, without another unlearning method on the same actions. This shows that FOMO suppresses actions, but leaves its advantage over existing methods unclear. The paper also does not report whether selecting motion-sensitive attention heads works better than using all attention heads. A comparison with the existing HunyuanVideo adaptation of ESD and an ablation without head selection would directly address these gaps.

Questions for authors

How often does FOMO remove the target action while keeping the subject’s appearance and unrelated background content unchanged? Please report the percentage of videos that meet both conditions on the motion test set, and compare with your HunyuanVideo adaptation of ESD using the same prompts and random seeds. Please explain how success is judged, allowing for pose or interaction changes needed to remove the action. A clear improvement over ESD on this measure would strengthen the paper’s central claim.

Does selecting motion-sensitive attention heads improve action removal or reduce unwanted scene changes? Please compare the current method with a version that uses all attention heads to construct the localization masks and compute the attention-output loss, keeping the remaining training settings fixed. Report both action removal and scene preservation. This would show whether the motion-specific design provides a measurable benefit.

5 comments

We thank the Reviewer for recognizing the value of action removal and the evidence of improved scene preservation. We also appreciate the acknowledgment that our ablation study of FOMO’s two objectives shows better preservation than ESD at the same Unsafe Rate. Following your suggestions, we conducted the additional experiments presented below.

W1/W2/Q1: We agree that full-frame DINO and LPIPS averages do not directly show how often action removal preserves the subject and background. As requested, we evaluated ESD adapted to HunyuanVideo on all six motions, using the same prompts and random seeds as FOMO. Inspired by the binary VLM evaluation in [1], we provide Qwen3-VL-8B-Instruct with six corresponding frame pairs from each original and unlearned video. The evaluator receives the intended action-removal instruction and answers two questions:

  1. Is the subject’s identity and overall appearance preserved, allowing for pose or interaction changes needed to remove the action?
  2. Is unrelated background content preserved, allowing for pose or interaction changes needed to remove the action?

Each criterion gets a yes/no answer and a short explanation. If it fails in any sampled frame, the score is 0. Joint Success means that the action is removed and both the subject and background are preserved in the same video. We check action removal using the same per-video labels used to calculate Motion Rate. We evaluate 20 videos for each of six actions (120 per method). As shown in Table A.1, FOMO removes the action while preserving both the subject and background in 50.83% of videos, compared with 5.00% for ESD (61 vs. 6 videos), despite similar Motion Rates (12.50% vs. 13.33%).

Following the evaluation protocol of HATIE [2], we compute CLIP, normalized DINO similarity, LPIPS, and MSE separately for subject and background regions, denoted by the prefixes s- and b-, respectively. Tables A.2 and A.3 show that FOMO performs better than ESD on average across all four metrics. For example: s-LPIPS is 0.1592 for FOMO vs. 0.2554 for ESD, and b-LPIPS is 0.0836 vs. 0.1774. We added the new evaluation results to the Appendix (Tables 19-21), along with the visual examples of Qwen’s evaluations (Figs. 12 and 13). We also show FOMO’s localization masks for cases with large scene changes (Fig. 11). These examples show whether the masks cover unrelated regions that may contribute to changes in the background.

Table A.1: Motion removal and VLM-based scene preservation across six actions.

MotionMethodMotion Rate (%) ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
RunningBase100————
ESD1552555
FOMO2060906045
FightingBase95————
ESD010401010
FOMO045554040
TongueBase100————
ESD305550
FOMO075757070
DancingBase100————
ESD1015301510
FOMO1065856560
JumpingBase85————
ESD05555
FOMO045604545
KnifeBase100————
ESD250500
FOMO4580757545
Avg. (6)Base96.67————
ESD13.336.6718.336.675.00
FOMO12.5061.6773.3359.1750.83

Table A.2: Subject preservation across six motions following the HATIE [2] protocol.

MotionMethods-CLIP ↑s-DINO ↑s-LPIPS ↓s-MSE ↓
RunningESD0.73900.75480.20630.0158
FOMO0.92740.94100.15970.0167
FightingESD0.89290.87390.25230.0208
FOMO0.90500.91640.18110.0162
TongueESD0.88360.86300.25600.0369
FOMO0.92890.93500.11480.0163
DancingESD0.89510.90130.25580.0336
FOMO0.92590.93540.17780.0179
JumpingESD0.86590.84220.29800.0223
FOMO0.90710.92160.17950.0196
KnifeESD0.89550.90020.26420.0336
FOMO0.94340.93780.14220.0144
Avg. (6)ESD0.86200.85590.25540.0272
FOMO0.92300.93120.15920.0169

Table A.3: Background preservation across six motions following the HATIE [2] protocol.

MotionMethodb-CLIP ↑b-DINO ↑b-LPIPS ↓b-MSE ↓
RunningESD0.92850.86500.21100.0421
FOMO0.97300.94750.07550.0130
FightingESD0.92450.90170.16650.0352
FOMO0.94650.91000.13620.0298
TongueESD0.94720.85690.09480.0150
FOMO0.97430.92450.03070.0059
DancingESD0.94530.87470.17350.0327
FOMO0.96360.92040.08870.0171
JumpingESD0.92000.77310.24390.0401
FOMO0.96370.92360.09110.0253
KnifeESD0.91030.84240.17490.0418
FOMO0.96990.94050.07920.0093
Avg. (6)ESD0.92930.85230.17740.0345
FOMO0.96520.92780.08360.0167

[1] Wang et al., “Rethinking Real Image Editing: Unleashing Diverse Editing Operators via Multi-Objective Optimization” WACV 2026.

[2] Ryu et al., “Towards Scalable Human-aligned Benchmark for Text-guided Image Editing” CVPR 2025.

W2/Q2: We thank the Reviewer for suggesting this ablation. We added Table A.4 comparing CHI-selected top-5, top-10, all heads, and random-5. Each head set is used for both mask construction and LHL_H, with all other settings, prompts, and random seeds fixed.

Top-5 performs best across all reported metrics. Compared with all heads, it lowers Motion Rate (12.50% vs. 15.83%) and improves Joint Success (50.83% vs. 44.17%). Random-5 performs worse despite using five heads, confirming that head choice matters. These results suggest that CHI selection focuses the intervention on the target motion while minimizing unintended changes to subject appearance and background. We also added this ablation to the main paper (see Table 8).

Table A.4: Effect of attention head selection on motion unlearning and scene preservation.

MethodMotion Rate (%) ↓DINO ↑LPIPS ↓Subject (%) ↑Background (%) ↑Both Preserved (%) ↑Joint Success (%) ↑
Base96.67——————
ESD13.330.42970.57146.6718.336.675.00
FOMO (top-5)12.500.70130.346761.6773.3359.1750.83
FOMO (all heads)15.830.65370.379853.3366.6751.6744.17
FOMO (random-5)25.830.60910.392247.5064.1745.8338.33
FOMO (top-10)16.670.63030.395050.0062.5047.5042.50

Dear Reviewer CS7p,

Thank you again for your constructive feedback. We hope our rebuttal addresses the points you raised. If you have any further questions, we’d be happy to answer them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.