- 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
- 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;
- 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;
- The experiments cover several tasks and datasets. I particularly appreciate the comparison between ( ) and ESD at the same Unsafe Rate in Table 6: the better scene similarity provides direct support for representation-level intervention.
Critical weaknesses
- 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;
- 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;
- 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;
- 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 ( ) 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
- Does FOMO retain its preservation advantage when non-target content is evaluated at comparable erasure rates?
- How well does the same adapter generate the safe motion and other retained actions after one action is erased?
- How does the source–safe feature difference limit the edit to the target attribute while preserving other attributes within the mask?
- 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
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:
-
Minimal prompt change: We change only the target phrase from to , 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.
-
Localized unlearning: We compute soft mask 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 . This improves joint subject and background preservation from 51.67% to 59.17% compared with using all heads (see Table 8 in paper).
-
Scene preservation: In Eq. (4), encourages the model to retain features from the original prompt at the first denoising step, with weight . Tokens with receive both the removal and preservation signals. For example, if removing the target concept slightly changes a person’s face, 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. guides the edit toward the replacement in , while 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 is empty, still anchors the scene to the original prompt , 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 .
Additionally, we evaluated several head configurations for motion unlearning (see Table C.3. below). Following W1/Q1, we asked Qwen two questions:
- Is the subject’s identity and overall appearance preserved, allowing for pose or interaction changes needed to remove the action?
- 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.
| Method | Motion Rate (%) ↓ | DINO ↑ | LPIPS ↓ | Subject (%) ↑ | Background (%) ↑ | Both Preserved (%) ↑ | Joint Success (%) ↑ |
|---|---|---|---|---|---|---|---|
| Base | 96.67 | — | — | — | — | — | — |
| ESD | 13.33 | 0.4297 | 0.5714 | 6.67 | 18.33 | 6.67 | 5.00 |
| FOMO (top-5) | 12.50 | 0.7013 | 0.3467 | 61.67 | 73.33 | 59.17 | 50.83 |
| FOMO (all heads) | 15.83 | 0.6537 | 0.3798 | 53.33 | 66.67 | 51.67 | 44.17 |
| FOMO (random-5) | 25.83 | 0.6091 | 0.3922 | 47.50 | 64.17 | 45.83 | 38.33 |
| FOMO (top-10) | 16.67 | 0.6303 | 0.3950 | 50.00 | 62.50 | 47.50 | 42.50 |
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!
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:
Is the subject’s identity and overall appearance preserved, allowing changes to clothing and skin exposure needed to remove nudity?
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) 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) 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 (LH) 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.
[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.