PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Łukasz Rudnik, Agnieszka Polowczyk, Alicja Polowczyk, Przemysław Spurek

arXiv:2609.39605v1cs.CV

TL;DR

Video unlearning has largely focused on static visual concepts, with motion and scene preservation receiving less attention. FOMO localizes and redirects target representations while preserving non-target content, including for motion unlearning. Across its evaluated tasks, it reports effective target suppression and a strong trade-off among erasure, scene preservation, and video quality.

  • Problem

    Video-unlearning methods largely focus on static visual concepts, while motion and preservation of surrounding scene content remain less explored.

    Source: §2 · p. 3

  • Method

    FOMO localizes concept-related representations, redirects them toward a safe alternative, and uses masked value preservation to limit changes to non-target content, including in motion unlearning.

    Source: §1 · p. 2, §6 · p. 9

  • Results

    Across unsafe-content, object, public-figure, and motion unlearning, FOMO reports effective target suppression while largely retaining scene structure and video quality, yielding a strong erasure–preservation trade-off.

    Source: §6 · p. 9

  • Takeaways & Limitations

    FOMO demonstrates selective video unlearning that addresses both visual concepts and motion while prioritizing preservation of the surrounding scene.

    Source: §6 · p. 9

    FOMO may require small contextual changes when unlearning a concept in a complex scene.

    Source: §5 · p. 9

Abstract

from arXiv · show

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. Code: https://github.com/gmum/FOMO Project Page https://gmum.github.io/FOMO

1 INTRODUCTION

Video unlearning must address temporal behavior as well as visual concepts, while preserving non-target scene content. FOMO combines targeted erasure, safe redirection, and scene preservation, and reports a strong trade-off across video-unlearning tasks.

1 INTRODUCTION

  • Video unlearning must handle temporal structure and motion in addition to spatial targets, while limiting changes to the surrounding scene.Small changes to a generation trajectory can propagate over time and alter scene evolution.Source: §1 · p. 2
  • Examples show FOMO removing running motion while largely preserving appearance and scene, and removing nudity while preserving dancing motion and the waterfall setting.The nudity example also illustrates replacing an unsafe visual concept without removing the accompanying motion.Source: Figure 2 · p. 2, Figure 3 · p. 2
  • FOMO localizes concept-related representations, redirects them toward a specified safe alternative, and preserves non-target scene information.Its two objectives address what to change and what to preserve.Source: §1 · p. 2
  • For motion unlearning, FOMO intervenes in motion-sensitive attention heads while aiming to retain the subject’s identity and surrounding scene.The method distinguishes motion targets from visual targets localized with frame-wise anchors.Source: §1 · pp. 2–3
  • FOMO reports the best trade-off among concept removal, scene preservation, and video quality across video-unlearning benchmarks.The evaluation covers unsafe content, objects, public figures, and motion.Source: §1 · pp. 2–3

2 RELATED WORKS

Prior video-unlearning work largely adapts image-focused approaches, leaving motion and scene preservation less explored. FOMO addresses both gaps by applying selective unlearning to visual concepts and motion while prioritizing scene preservation.

2 RELATED WORKS

  • Most concept-unlearning methods target text-to-image models, while video methods include inference-time guidance and model fine-tuning or weight updates.The related work names SAFREE and VideoEraser as inference-time approaches and T2VUnlearning, EraseAnything++, NullSCE, and Low-Rank Refusal Vector as model-modification approaches.Source: §2 · p. 3
  • Existing video unlearning largely follows image-unlearning settings focused on visual targets, while motion and temporal behaviors remain less explored.Prior 3D motion-unlearning work does not cover appearance or surrounding scene context within its generative model.Source: §2 · p. 3
  • FOMO uses source and safe prompts to define the target and its counterpart, then localizes target representations and preserves the rest of the scene.The overview illustrates nudity being replaced with a pink dress while largely retaining the person and scene.Source: Figure 5 · p. 4
  • Scene preservation remains insufficiently explored as a primary objective across motion and video unlearning, a gap FOMO addresses alongside motion coverage.FOMO applies scene-preserving unlearning to both visual concepts and motion in generative video models.Source: §2 · pp. 3–4

3 METHODOLOGY

FOMO localizes concept-related video representations, redirects them toward a safe prompt, and preserves non-target scene information. It extends this selective approach to motion by focusing localization and unlearning on temporally sensitive attention heads.

3.1 WHERE IS THE CONCEPT PRESENT?

  • For visual concepts, FOMO combines a concept-associated video-token anchor in each frame with representational similarities to create a localization mask.The mask is applied independently by block and head, and masks are stacked across frames.Source: §3.1 · p. 5

3.2 HOW TO UNLEARN THE CONCEPT?

  • A frozen base model supplies source and safe-prompt representations, and a concept mask confines the student’s shift toward the safe representation to target-associated regions.Only LoRA parameters are optimized for representation alignment while the base parameters remain frozen.Source: §3.2 · p. 5

3.3 WHAT SHOULD BE PRESERVED?

  • FOMO preserves scene structure by matching value representations in non-target regions, applying this loss at high-noise timesteps.The preservation objective’s strength is controlled by λV.Source: §3.3 · p. 6

3.4 MOTION-SELECTIVE UNLEARNING

  • FOMO identifies motion-sensitive heads using frame-wise variation, then uses their similarity maps for localization and evaluates the unlearning loss only on those heads.Heads are ranked by the Calinski–Harabasz index; higher scores indicate more distinct representations across frames.Source: §3.4 · p. 6

4 EXPERIMENTS

FOMO is evaluated across unsafe-content, object, public-figure, and motion unlearning against video-unlearning baselines. Results report target suppression alongside scene preservation and video quality.

  • Evaluation spans unsafe-content, object, public-figure, and motion tasks, comparing FOMO with inference-time methods, T2VUnlearning, and adapted ESD.Source: §4 · p. 7
  • For nudity, FOMO lowers Unsafe Rate from 53.14% to 19.54% and achieves the best scene preservation; for other harms, the average falls from 59.06% to 12.58% with higher scene similarity than SAFREE.Source: §4 · p. 8
  • 92.41 ESR-1 and 77.24 ESR-5: FOMO achieves substantially higher object erasure than Negative Prompt and SAFREE while keeping preservation close to the base model.Source: Table 3 · p. 8, §4 · p. 8
  • 0.5793 to 0.0822 average ID similarity: FOMO suppresses the erased public figure while retaining 0.5000 average similarity for the other identities and largely preserving pose and background.Source: §4 · p. 8, Figure 10 · p. 8
  • 96.67% to 12.50% average Motion Rate: FOMO reduces target motion while maintaining high temporal smoothness and scene similarity.Source: Table 5 · p. 9, §4 · p. 9

5 ABLATION STUDY

The ablations compare FOMO’s erasure and preservation objectives, including masked versus global preservation. The combined objectives give the strongest reported erasure–preservation balance, while complex scenes can require contextual changes.

  • FOMO scores 80.04 on VBench Final Score and has the highest DINO similarity, 0.739, and lowest LPIPS, 0.300, among unlearning methods.Source: Table 4 · p. 9, §5 · p. 9
  • 4% Unsafe Rate: LH alone reaches ESD-level erasure with substantially better scene preservation, while the masked-preservation term alone leaves generation unchanged.Source: §5 · p. 9
  • 26% Unsafe Rate: combining erasure with masked preservation yields the best erasure–preservation balance; applying preservation across the whole video raises Unsafe Rate to 44%.Source: §5 · p. 9
  • Unlearning a concept in a complex scene may require small contextual changes despite the scene-preservation objective.Source: §5 · p. 9

6 CONCLUSION

FOMO modifies concept-related representations while using masked value preservation to limit changes to non-target content, and targets motion-sensitive attention heads for motion unlearning. The conclusion reports effective suppression with scene structure and video quality largely retained, while noting that complex unsafe scenes may need contextual changes.

6 CONCLUSION

  • FOMO combines concept-related representation modification with masked value preservation, and focuses motion unlearning on motion-sensitive attention heads.Source: §6 · p. 9

APPENDIX

  • The supplementary material covers implementation settings, tasks and evaluation protocols, detailed results, prompt construction and ablations, scene-change cases, and qualitative visualizations.Source: APPENDIX · p. 14

A IMPLEMENTATION DETAILS

  • On HunyuanVideo, masks use 13 selected double-stream blocks; motion retains five heads, concept erasure aggregates heads, and training updates rank-8 query/key LoRAs only.Source: A IMPLEMENTATION DETAILS · p. 14

B.1 GENERAL SAFETY UNLEARNING

  • Safety evaluation uses GEN, Ring-A-Bell, SafeSora, and T2VSafetyBench, with Qwen3-VL assessing unsafe frames and LPIPS, DINO, and VBench assessing preservation and video quality.Source: §B.1 · pp. 14–15

B.2 OBJECT UNLEARNING

  • Object erasure tests ten Imagenette classes with safe replacements; ESR-k is 1 − Top-k accuracy on the erased class, while PSR-k averages Top-k accuracy over the other nine.Source: §B.2 · p. 16

B.3 PUBLIC FIGURE UNLEARNING

  • Public-figure evaluation trains separate adapters for five identities and uses ArcFace similarity to measure erased-identity suppression and preservation of the other four.Source: §B.3 · p. 16

B.4 MOTION UNLEARNING

  • Motion evaluation covers six dynamic concepts, measures target behavior with Qwen3-VL, and assesses scene and video preservation with DINO, LPIPS, and VBench.Source: §B.4 · p. 17, Table 7 · p. 17

C ADDITIONAL EVALUATION RESULTS OF ERASING THE GENERAL SAFETY

Across four nudity-unlearning benchmarks, FOMO combines low Unsafe Rate with strong preservation of the original generations. It also retains broad video quality and scene structure on VBench, while competing methods expose a trade-off between erasure and preservation.

C ADDITIONAL EVALUATION RESULTS OF ERASING THE GENERAL SAFETY

  • FOMO maintains a low Unsafe Rate alongside strong DINO/LPIPS preservation across four nudity-unlearning benchmarks.The reported scores indicate fidelity to base-model generations while balancing concept removal with preservation of the surrounding scene.Source: C ADDITIONAL EVALUATION RESULTS OF… · p. 18, Table 8 · p. 18, Table 9 · p. 18
  • ESD achieves a very low Unsafe Rate but much worse preservation, whereas Negative Prompt preserves content relatively well without improving Unsafe Rate.Negative Prompt’s Unsafe Rate can even increase; the passage says classifier-free guidance away from the unsafe prompt is not enough for effective unlearning.Source: C ADDITIONAL EVALUATION RESULTS OF… · p. 18
  • FOMO scores best on DINO and LPIPS among compared unlearning methods and retains a Final Score of 80.04 versus 81.05 for the base model.Under custom VBench inputs, results also remain close to the base model across consistency, motion, aesthetic, and imaging measures.Source: C ADDITIONAL EVALUATION RESULTS OF… · p. 18, Table 12 · p. 19

D ADDITIONAL EVALUATION RESULTS OF ERASING THE IMAGENETTE CLASSES

On Imagenette, FOMO achieves the highest reported object-erasure scores while keeping preservation close to the base model.

D ADDITIONAL EVALUATION RESULTS OF ERASING THE IMAGENETTE CLASSES

E ADDITIONAL EVALUATION RESULTS OF ERASING THE PUBLIC FIGURE

For public-figure unlearning on HunyuanVideo, FOMO reduces similarity to erased identities while preserving similarity to remaining identities.

E ADDITIONAL EVALUATION RESULTS OF ERASING THE PUBLIC FIGURE

F ADDITIONAL EVALUATION RESULTS OF ERASING THE MOTION

FOMO reduces source-motion behavior across dynamic concepts, with reported Dynamic Degree changes for running, jumping, and dancing. Pair-prompt tests find nudity unlearning only mildly sensitive to prompt choice, while object mappings expose an erasure–preservation trade-off.

F ADDITIONAL EVALUATION RESULTS OF ERASING THE MOTION

  • For running→walking, Dynamic Degree falls from 0.9 to 0.7; for jumping→standing still it falls from 0.85 to 0.1, and for dancing→standing still from 0.75 to 0.3.Across six selected dynamic concepts, Source Motion Rate decreases substantially, particularly for jumping, fighting, and sticking out the tongue.Source: F ADDITIONAL EVALUATION RESULTS OF… · p. 21, Table 15 · p. 21

G EFFECT OF PAIR PROMPT SELECTION

  • Across nudity variants with indoor or outdoor scenes, source→safe prompt-pair choice has only a minor effect on unlearning.Each source prompt is paired with a corresponding clothed version while preserving the remaining context; Table 16 averages Unsafe Rate, DINO similarity, and LPIPS across four benchmarks.Source: G EFFECT OF PAIR PROMPT SELECTION · p. 22, Table 16 · p. 22
  • Object unlearning remains effective across safe mappings, but the null mapping yields strongest erasure at the cost of reduced preservation.The results use Safe Mapping 2 in the main paper; Table 17 reports per-class results under different mappings.Source: G EFFECT OF PAIR PROMPT SELECTION · p. 22, Table 17 · p. 22
  • Safe Mappings 1 and 2 share scene structure, while a null prompt can redirect a parachute toward a boat mismatched to the aerial scene without restoring the erased class.The mapping choices are listed in Table 18, and Figure 11 compares object unlearning with different safe mappings.Source: G EFFECT OF PAIR PROMPT SELECTION · p. 22, Figure 11 · p. 23, Table 18 · p. 24

H SCENE PRESERVATION IN CHALLENGING CASES

Removing a target concept can require changing the surrounding scene when its original behavior or arrangement depends on that concept. FOMO therefore aims to preserve non-target content while allowing changes needed for effective, coherent unlearning.

H SCENE PRESERVATION IN CHALLENGING CASES

  • Exact spatial, compositional, or pixel-level reconstruction is not always compatible with removing concepts embedded in a scene.The authors ask what should be preserved when elements of the original scene themselves form part of the target concept.Source: H SCENE PRESERVATION IN CHALLENGING… · p. 24
  • After removing a dead body, people and police no longer focus on its former location, avoiding an unnatural crowd reacting to an empty space.The example illustrates why preserving the original behavior alongside the removed target may be semantically inconsistent.Source: H SCENE PRESERVATION IN CHALLENGING… · p. 24, Figure 12 · p. 24
  • After an explosion and its consequences disappear, people running while laughing may produce a more coherent scene than retaining panic without the violent event.The authors note that behavioral and contextual changes may be necessary for semantic coherence.Source: H SCENE PRESERVATION IN CHALLENGING… · p. 24, Figure 13 · p. 25
  • Removing burglary semantics changes a break-in scene into a normal interaction when a woman greets the man entering the house.The added person changes the scene and gives it a more positive interpretation.Source: H SCENE PRESERVATION IN CHALLENGING… · p. 24, Figure 14 · p. 25
  • Scene preservation may require changing surrounding behavior or arrangement when those elements are closely tied to the concept being removed.The authors aim to retain as much non-target content and scene structure as possible while allowing such changes.Source: H SCENE PRESERVATION IN CHALLENGING… · p. 24

I ADDITIONAL QUALITATIVE RESULT

The appendix adds qualitative examples and comparisons across safety, nudity, public-figure, motion, and brand unlearning. Captions describe corrected outputs or the intended comparison, while some examples specifically report scene preservation alongside concept removal.

I ADDITIONAL QUALITATIVE RESULT

  • Additional qualitative results cover general safety, nudity, public figures, motion concepts, and brand unlearning.Table 19 summarizes experiment types, settings, and figure references for the appendix examples.Source: I ADDITIONAL QUALITATIVE RESULT · p. 26, Table 19 · p. 26
  • The nudity examples provide qualitative comparisons across methods on GEN, Ring-A-Bell, and T2VSafetyBench.The captions identify HunyuanVideo as the model used for these comparisons.Source: Figure 15 · p. 26, Figure 16 · p. 27, Figure 17 · p. 27
  • For gore on copyrighted children’s characters, FOMO suppresses gore while largely preserving the surrounding scene and characters’ overall appearance.The appendix also shows corrected versions for gore and violence, with uncensored frames above corrected frames.Source: Figure 18 · p. 28, Figure 19 · p. 29, Figure 20 · p. 30
  • The public-figure examples remove one identity at a time, with the blue frame marking the identity to forget and other faces intended to remain unchanged.The caption describes the original model in the leftmost column and the identity-removed models in the remaining columns.Source: Figure 22 · p. 31
  • Motion examples cover brandishing a knife, sticking out the tongue, dancing, fighting, jumping, and running.The captions present uncensored frames above corrected frames for these HunyuanVideo examples.Source: Figure 23 · p. 32, Figure 24 · p. 33, Figure 25 · p. 34
  • Sony, Starbucks, Ferrari, and Louis Vuitton products remain visible after unlearning while their brand identities are removed.The examples cover Sony and Starbucks, and Ferrari and Louis Vuitton.Source: Figure 35 · p. 44, Figure 36 · p. 44

Related papers

Loading 2609.39605v1…