Pith. sign in

REVIEW 4 major objections 5 minor 46 references

A training-free method, SIRUS, claims to erase target concepts during video generation — no weight updates — with the best forgetting-versus-quality balance among tested baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 03:03 UTC pith:BS6P7L7V

load-bearing objection A useful new training-free T2V unlearning recipe and a genuinely multi-axis evaluation framework, but the headline numbers are undercut by missing prompt-deletion/no-op controls that could explain much of the effect. the 4 major comments →

arxiv 2607.14194 v1 pith:BS6P7L7V submitted 2026-07-15 cs.CV cs.LG

Inference-Time Concept Suppression and Video-Centric Evaluation for Text-to-Video Models

Wenxuan Chen , Wenjie Feng This is my paper
classification cs.CV cs.LG
keywords concept unlearningtext-to-video generationdiffusion modelsinference-time interventionconcept subspaceresidual guidancemachine unlearning evaluationvideo generation safety
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper is trying to establish that concept-level unlearning in text-to-video generation can be done at inference time, with no parameter updates: given textual aliases of a target concept, SIRUS locates the concept inside a prompt and steers the diffusion trajectory away from it during sampling. The paper's core claim is that this localization-versus-suppression split — a text-embedding subspace built from aliases, a hierarchical trigger that marks target tokens, and a scheduled residual subtracted from the guidance output — removes safety, object, and style concepts more reliably than prior inference-time baselines while degrading video quality far less. It also argues that T2V unlearning must be judged at the video level, not on isolated frames, and provides an evaluation framework that reports forgetting, preservation, video quality, robustness, and efficiency separately. If the central claim is right, the practical consequence is that one frozen video generator can serve many forget policies by swapping an alias set and a few profile parameters, at the cost of roughly a quarter more generation time, and the recipe transfers to a second backbone.

Core claim

The central discovery is that the hard part of T2V unlearning is not how strongly to intervene but where the concept lives in the prompt and what direction would restore it. SIRUS encodes templated alias phrases through the frozen text encoder and takes the singular directions of the centered alias embeddings to form an orthonormal concept subspace B. A hierarchical trigger — exact alias match, then subspace similarity, then a conservative fallback — marks which tokens carry target evidence, and only those tokens are projected away from B. At sampling time, a positive concept-reference condition is mixed from the most similar reference prompts, and the denoiser's prediction under it versus t

What carries the argument

The load-bearing object is the concept subspace: an orthonormal row-basis B built by encoding a small bank of templated alias prompts with the frozen text encoder, centering the alias token embeddings, and keeping their top singular directions. The same projection operator does two jobs — measuring per-token similarity for the trigger rule and removing the concept-aligned component from triggered tokens. The second mechanism is the concept residual d_t = ε_c − ε_t, the difference between the denoiser's output under a positive concept-reference condition and under the edited prompt condition, read as the direction that would restore the target, so a controlled subtraction steers generation aw

Load-bearing premise

The load-bearing premise is that the linear subspace built from a small alias-template bank (Section 3.1) reliably spans the target concept across arbitrary prompts, and that the residual d_t = ε_c − ε_t (Section 3.4) faithfully points at what would restore it, so subtracting a capped, momentum-smoothed version suppresses the concept without collateral damage — a premise tested on only five concepts, with profile hyperparameters never reported and the paper itself conceding (

What would settle it

Take a concept outside the five tested (say 'bicycle') and run SIRUS over a few hundred prompts on the same backbone, scoring video-level forgetting with the same classifier. If success collapses toward the no-intervention baseline, or removing the target also distorts non-target content (large LPIPS jump or object-recall drop), the subspace-faithfulness premise fails. Sharper: vary the alias bank size from one to about twenty on a fixed concept — monotone improvement with more aliases confirms the subspace carries the method, while flat or erratic success means something else does. Or disable

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single frozen video generator can enforce multiple forget policies by swapping the alias set and profile, without maintaining separately edited model copies; the measured cost is about a 24% increase in per-video generation time on the primary backbone, not offline retraining.
  • The video-oriented evaluation framework exposes a failure mode that frame-level checks miss: an aggressive baseline reaches high forgetting on nudity by erasing the person itself (person retention 33.4% versus 78.5% for SIRUS), so forgetting and preservation must be read jointly.
  • The component ablation attributes the largest share of forgetting to the sampling-side concept-reference branch — nudity success drops from 80.0% to 42.0% without it — implying prompt editing alone would not carry the method.
  • The same recipe transfers to a second, larger backbone with 73.6% average forgetting success, suggesting the approach is not tied to one model's text-embedding geometry.
  • The paper itself concedes a boundary: visually salient, temporally persistent targets such as parachute keep residual traces, so inference-time suppression is not uniform across concept types.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The machinery is almost entirely concept-generic: subspace localization plus residual subtraction would apply to other generative domains (image, audio, 3D) where a concept can be named by aliases; only the temporal schedule is video-specific, and it could be reinterpreted as a per-timestep intervention profile.
  • The paper does not vary the alias bank size, template phrasing, or profile parameters; the sharpest unstated test is how forgetting success degrades as aliases become sparser or more figurative, since the entire method rests on the subspace spanning the concept's visual expression.
  • The per-concept pattern — strong on compact objects like garbage truck, weak on global, persistent targets like parachute — suggests the optimal intervention profile differs by concept type, and a per-concept profile search is a natural extension the paper does not attempt.
  • Because suppression happens in text-embedding space, one testable hypothesis is that paraphrase and obfuscation jailbreaks that preserve semantic proximity to the alias subspace are deflected, while attacks that route around the embedding entirely may not be — a distinction the robustness study does not yet separate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SIRUS, a training-free inference-time method for concept unlearning in text-to-video diffusion models. SIRUS builds a contextualized concept subspace from textual aliases, localizes target-related prompt evidence via a hierarchical trigger, projects triggered token embeddings away from the subspace, retrieves a positive concept-reference branch, and subtracts a capped, momentum-smoothed, time-scheduled residual during classifier-free guidance sampling. The paper also introduces VUEF, a video-oriented evaluation framework with separate forgetting, preservation, quality, robustness, and efficiency axes. On CogVideoX across five safety/object/style concepts, SIRUS is reported to reach 70.4% average forgetting success and 25.7% frame hit versus 44.4%/47.2% for VideoEraser, while reducing the VBench quality drop from -0.043 to -0.016. A Wan2.2 transfer study and several ablations are also presented.

Significance. If confirmed, the paper would make a useful contribution: an inference-time, training-free unlearning mechanism that separates concept localization from suppression, together with a multi-axis T2V evaluation protocol. The per-concept tables, the nudity-specific person-retention analysis, and the component ablation are informative, and the authors are transparent about the parachute difficulty, the preservation/forgetting trade-off, and the Refusal Vector backbone mismatch. The external MultiClf classifier and VBench metrics reduce circularity. However, the central trade-off claim currently rests on missing experimental controls and unreported hyperparameters, so the significance of the headline numbers is conditional.

major comments (4)
  1. [§4, Table 1; §3.2] The central forgetting comparison lacks a no-intervention baseline and a prompt-deletion control. Table 1 shows no row for the unmodified CogVideoX on the same prompts, so the original model's Fail/AnyHit/FrameHit rates are unknown. Because SIRUS's first trigger level is exact alias/anchor matching and its text-side edit is a local projection of triggered tokens, and because the evaluation prompts are adapted from UnlearnDiffAtk and explicitly name the target concept, a trivial control that deletes or masks the target words is a natural competitor and could explain much of the observed forgetting and quality preservation. Without such a control, the 70.4% vs 44.4% gap is not attributable to the subspace/residual mechanism.
  2. [§3.2–§3.4] The method is not reproducible from the manuscript. The profile-dependent coefficients απ, γπ, ηπ, ρπ, βπ, and r_{e,π} are defined but their numerical values or the mapping from evidence profiles to parameter sets are not reported. The same holds for the subspace rank K/energy threshold, similarity threshold, relaxed-fallback token budget, and top-k reference count. These parameters determine the intervention strength and schedule; without them the reported trade-off cannot be independently reconstructed or compared across methods.
  3. [§5.1, Tables 1–3] The experimental tables report point estimates without sample sizes, seed variation, or confidence intervals. N=16 is the number of sampled frames per video, not the number of videos per concept. Several comparisons are modest (church 48 vs 34; parachute 50 vs 28), and the average ΔQ difference (-0.016 vs -0.043) has no error bar. The 'strongest forgetting-quality trade-off' claim is not statistically supported. Report the number of prompts/videos per concept and at least three seeds or confidence intervals.
  4. [§5.2.1, §B.1] The primary 'success rate' uses a permissive failure threshold: with N=16 sampled frames, a video with three or fewer target-classified frames counts as successfully unlearned. For safety concepts, this conflates residue reduction with forgetting. The choice K=4 is asserted without sensitivity analysis. The appendix reports AnyHit, but the main-text headline still depends on K. Report success for K=0 and at least one intermediate threshold, or justify K=4 more rigorously.
minor comments (5)
  1. [§3.4, §3.3] Copy-editing needed: §3.4 contains the duplicated fragment 'Finally, SIRUS applies a three-part temporal schedule. Third, SIRUS applies a three-part temporal schedule.' and §3.3 has 'the trigger mask is used to from a positive reference prompt bank.'
  2. [§3.1, §4] The symbol K is used both for the concept-subspace rank and for the forgetting failure threshold. Rename one of them (e.g., K_sub and K_fail) to avoid confusion.
  3. [Abstract; §5.1] The comparison to Refusal Vector is not backbone-controlled. The paper acknowledges this in §5.5, but the abstract's 'fully evaluated baselines' phrasing should make explicit that only VideoEraser is a same-backbone full baseline on all five concepts.
  4. [§5.4, Table 5] The Wan2.2 generalization claim is supported in the main text only by forgetting metrics; preservation and VBench quality appear only in the appendix. Add a one-line main-text summary or a Wan2.2 baseline row.
  5. [§4] The anonymized VUEF implementation is released, but no SIRUS implementation link is given. For an inference-time method, releasing the inference wrapper would materially aid reproducibility.

Circularity Check

0 steps flagged

No significant circularity: SIRUS is an empirical inference-time intervention whose headline numbers are measured by external detectors/benchmarks, not entailed by its definitions.

full rationale

No load-bearing circular step is present. The method constructs a concept subspace B from textual aliases (Sec. 3.1) and uses it for triggering, projection, and residual subtraction (Secs. 3.2-3.4); this is a constructive heuristic, not a derivation that presupposes the measured outcome. The forgetting results are computed with the 'well-trained MultiClf evaluator from IGMU [21]' (Sec. 4), a pre-existing classifier used outside this paper's fitted values, and video quality is measured with external VBench metrics. The self-reference to the authors' IGMU/EvalIGMU work [21] (Wenjie Feng is a co-author of both) is a minor evaluation-stack inheritance, but it does not reduce the results to the paper's inputs: MultiClf and VBench are independent/external and would yield unfavorable numbers if SIRUS failed. The paper also states its own limitations, e.g., parachute-like persistent targets remain difficult and Refusal Vector is not backbone-controlled. The absence of a no-op/prompt-deletion control and unreported hyperparameters (alpha_pi, gamma_pi, eta_pi, etc.) are evaluation-control and reproducibility gaps, not circularity under the stated criteria. No equation is identical to another by construction, and no fitted parameter is renamed as a prediction. Score 1 reflects only the mild self-citation in the evaluation framework; no circular reduction was found.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 0 invented entities

The concept subspace and reference branch are internal representations, not independently testable entities. No new physical or external entities are postulated; the method's load-bearing assumptions are the text-embedding-to-visual-concept alignment and the residual direction interpretation, both unproven empirically beyond the reported experiments.

free parameters (11)
  • Subspace rank K / energy threshold
    Rank of the concept subspace B; not reported; affects triggering and projection strength.
  • Prompt projection coefficient α_π
    Profile-dependent strength of projection; no values given; directly controls how much target evidence is removed.
  • Reference mixing coefficient γ_π
    Mixes unconditional condition c_u with concept-reference condition c_c; not reported.
  • Erase guidance scale η_π
    Scales the subtracted concept residual; no value reported; central to suppression strength.
  • Residual cap ρ_π
    Caps residual magnitude relative to text residual; not reported.
  • Momentum β_π
    Momentum smoothing coefficient; not reported.
  • Erase start/end ratios r_s, r_{e,π}
    Temporal schedule window; not reported; controls late-step interference.
  • Similarity threshold and relaxed fallback budget
    Determines trigger mask; not reported.
  • Top-k reference count
    Number of reference prompts aggregated; not reported.
  • Alias sets and template bank
    Exact alias lists for five concepts and template prompts are not given; they are inputs to the subspace construction.
  • Failure threshold K and sample frames N = K=4, N=16
    Evaluation thresholds chosen; affect reported success/hit rates.
axioms (5)
  • domain assumption The text encoder embeddings of contextualized aliases capture concept-relevant visual semantics (Section 3.1).
    The method builds a subspace from prompt embeddings and assumes it aligns with the visual concept to be erased; no evidence given that this holds across diverse prompts or backbones.
  • domain assumption The concept residual d_t = ε_c − ε_t approximates the direction that restores the target concept, so subtracting it suppresses the concept (Section 3.4).
    This is the core operational premise of residual unlearning; it is asserted without formal justification.
  • domain assumption The MultiClf evaluator (from IGMU [21]) correctly classifies target concept presence in generated video frames (Section 4).
    All forgetting metrics depend on this external classifier; its accuracy on T2V frames is not reported.
  • standard math Classifier-free guidance prediction ε_base = ε_u + s(ε_t − ε_u) is a valid base; combining it with a subtracted residual yields valid denoising (Section 3.4).
    Standard CFG; subtraction of a residual is a heuristic extension.
  • ad hoc to paper The intervention does not break the denoiser's stability across the temporal schedule (Section 3.4).
    The cosine schedule and residual cap are introduced to avoid late-step interference, but no stability analysis is offered.

pith-pipeline@v1.3.0-alltime-deepseek · 21437 in / 15270 out tokens · 128436 ms · 2026-08-02T03:03:41.574351+00:00 · methodology

0 comments
read the original abstract

Text-to-video (T2V) generators can synthesize realistic and temporally coherent videos, but controllably removing a target concept from a generator remains difficult. Unlike text-to-image concept erasure, T2V unlearning must suppress a target concept that may persist across frames while preserving non-target subjects, actions, scenes, and temporal structure. We propose \textbf{SIRUS}, a training-free inference-time framework for concept-level T2V unlearning. Given textual aliases of a target concept, SIRUS localizes target-related prompt evidence and suppresses target expression during sampling, without updating the text encoder or denoising network. We further introduce a video-oriented evaluation framework for T2V unlearning that separately measures target forgetting, non-target preservation, video quality, jailbreak robustness, and efficiency, using video-level failure criteria, frame-level residue statistics, paired preservation analysis, VBench-based quality diagnostics, and deployment overhead measurement. Across five safety, object, and style concepts on CogVideoX, SIRUS reaches 70.4\% average forgetting success and 25.7\% average frame hit, compared with 44.4\% / 47.2\% for VideoEraser, while reducing the average VBench quality drop from -0.043 to -0.016, yielding the strongest forgetting-quality trade-off among fully evaluated baselines. Transfer experiments on Wan2.2 further suggest that SIRUS generalizes across modern T2V backbones.

Figures

Figures reproduced from arXiv: 2607.14194 by Wenjie Feng, Wenxuan Chen.

Figure 1
Figure 1. Figure 1: Teaser examples of text-to-video unlearning on three representative prompts: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SIRUS. Starting from textual aliases of the target concept, SIRUS builds [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative overview on four representative concepts. In each panel, the baseline contains [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Supplementary nudity examples. SIRUS removes the unsafe nudity concept while [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: More supplementary nudity examples. These additional cases further illustrate that [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Supplementary church examples. Across four prompts, SIRUS suppresses church-defining [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Supplementary garbage-truck examples. Across four prompts, SIRUS suppresses the [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Supplementary parachute examples. Across four prompts, parachute remains a challenging [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Supplementary Van Gogh-style examples. Across four prompts, SIRUS weakens the target [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 5 canonical work pages

  1. [1]

    Abubakar Abid, Martin Jinye Zhang, Vivek Kumar Bagaria, and James Y. Zou. Explor- ing patterns enriched in a dataset with contrastive principal component analysis.Nature Communications, 9, 2018

    2018

  2. [3]

    Pappas, Florian Tramèr, Hamed Hassani, and Eric Wong

    Patrick Chao, Edoardo Debenedetti, Alexander Robey, Maksym Andriushchenko, Francesco Croce, Vikash Sehwag, Edgar Dobriban, Nicolas Flammarion, George J. Pappas, Florian Tramèr, Hamed Hassani, and Eric Wong. JailbreakBench: An open robustness benchmark for jailbreaking large language models. InAdvances in Neural Information Processing Systems, 2024

    2024

  3. [4]

    Video unlearn- ing via low-rank refusal vector

    Simone Facchiano, Stefano Saravalle, Matteo Migliarini, Edoardo De Matteis, Alessio Sampieri, Andrea Pilzer, Emanuele Rodolà, Indro Spinelli, Luca Franco, and Fabio Galasso. Video unlearn- ing via low-rank refusal vector. InThe Fourteenth International Conference on Learning Rep- resentations. OpenReview.net, 2026. URLhttps://openreview.net/forum?id=U1XBHtXl7Y

    2026

  4. [5]

    Robert M. French. Catastrophic forgetting in connectionist networks.Trends in Cognitive Sciences, 3(4):128–135, 1999

    1999

  5. [7]

    Unified concept editing in diffusion models

    Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzynska, and David Bau. Unified concept editing in diffusion models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5099–5108, 2024. doi: 10.1109/WACV57701.2024. 00503. URLhttps://doi.org/10.1109/WACV57701.2024.00503

    arXiv 2024

  6. [8]

    Reliable and efficient concept erasure of text-to-image diffusion models

    Chao Gong, Kai Chen, Zhipeng Wei, Jingjing Chen, and Yu-Gang Jiang. Reliable and efficient concept erasure of text-to-image diffusion models. InProceedings of the European Conference on Computer Vision, pages 73–88, 2024. doi: 10.1007/978-3-031-73668-1_5. URL https://doi.org/10.1007/978-3-031-73668-1_5

    work page doi:10.1007/978-3-031-73668-1_5 2024

  7. [9]

    AnimateDiff: Animate your personalized text-to-image diffu- sion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffu- sion models without specific tuning. InThe Twelfth International Conference on Learning Rep- resentations. OpenReview.net, 2024. URLhttps://openreview.net/forum?id=Fx2SbBgcte

    2024

  8. [10]

    CLIPScore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514–7528. Association for Computational Linguistics, 2021. doi: 10.18653/v1/2021.emnlp-main.595. URLhttps://doi. org/1...

    doi:10.18653/v1/2021.emnlp-main.595 2021

  9. [11]

    GANs trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. InAdvances in Neural Information Processing Systems, volume 30, pages 6626–6637, 2017. URL https://proceedings.neurips.cc/paper/2017/hash/ 8a1d694707eb0fefe65871369074926d-Abstract.html

    2017

  10. [12]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. doi: 10.48550/arXiv.2207.12598. URLhttps://doi.org/10.48550/ arXiv.2207.12598

    Pith/arXiv arXiv doi:10.48550/arxiv.2207.12598 2022

  11. [13]

    Gritsenko, Diederik P

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303,

    Pith/arXiv arXiv

  12. [14]

    ConceptVoid: Precision multi- concept erasure in generative video diffusion.Mathematics, 13(16):2652, 2025

    Zhongbin Huang, Xingjia Jin, Cunkang Wu, and Wei Mao. ConceptVoid: Precision multi- concept erasure in generative video diffusion.Mathematics, 13(16):2652, 2025. doi: 10.3390/ math13162652. URLhttps://doi.org/10.3390/math13162652

    work page doi:10.3390/math13162652 2025

  13. [16]

    Ultralytics YOLOv8.https://github.com/ ultralytics/ultralytics, 2023

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8.https://github.com/ ultralytics/ultralytics, 2023. Software, version 8.0.0

    2023

  14. [17]

    HunyuanVideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.03603 2024

  15. [18]

    Ablating concepts in text-to-image diffusion models

    Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634–22645. IEEE, 2023. doi: 10.1109/ ICCV51070.2023.02074. URLhttps://doi.org/10.1109/ICCV51070.2023.02074

    arXiv 2023

  16. [19]

    Learning blind video temporal consistency

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. InProceedings of the European Conference on Computer Vision, pages 179–195, 2018. doi: 10.1007/978-3-030-01267-0_11. URLhttps: //doi.org/10.1007/978-3-030-01267-0_11

    work page doi:10.1007/978-3-030-01267-0_11 2018

  17. [20]

    Fréchet Video Motion Distance: A metric for evaluating motion consistency in videos.arXiv preprint arXiv:2407.16124,

    Jiahe Liu, Youran Qu, Qi Yan, Xiaohui Zeng, Lele Wang, and Renjie Liao. Fréchet Video Motion Distance: A metric for evaluating motion consistency in videos.arXiv preprint arXiv:2407.16124,

    Pith/arXiv arXiv

  18. [21]

    Rethinking machine unlearning in image generation models

    Renyang Liu, Wenjie Feng, Tianwei Zhang, Wei Zhou, Xueqi Cheng, and See-Kiong Ng. Rethinking machine unlearning in image generation models. InProceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security, pages 993–1007, 2025. doi: 10.1145/3719027.3744793. URLhttps://doi.org/10.1145/3719027.3744793

    arXiv 2025

  19. [22]

    AutoDAN: Generating stealthy jailbreak prompts on aligned large language models

    Xiaogeng Liu, Nan Xu, Muhao Chen, and Chaowei Xiao. AutoDAN: Generating stealthy jailbreak prompts on aligned large language models. InThe Twelfth International Conference on Learning Representations. OpenReview.net, 2024. URLhttps://openreview.net/forum? id=7Jwpw4qKkb

    2024

  20. [24]

    FETV: A benchmark for fine-grained evaluation of open-domain text-to-video generation

    Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. FETV: A benchmark for fine-grained evaluation of open-domain text-to-video generation. InAdvances in Neural Information Process- ing Systems, 2023. URL https://papers.nips.cc/paper_files/paper/2023/hash/ c481049f7410f38e788f67c171c64ad5-Abstract-Datasets_and_Bench...

    2023

  21. [25]

    MACE: Mass concept erasure in diffusion models

    Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. MACE: Mass concept erasure in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430–6440. IEEE, 2024. doi: 10.1109/CVPR52733.2024.00615. URLhttps://doi.org/10.1109/CVPR52733.2024.00615

    arXiv 2024

  22. [26]

    T2VSafetyBench: Evaluating the safety of text-to-video generative models

    Yibo Miao, Yifan Zhu, Yinpeng Dong, Lijia Yu, Jun Zhu, and Xiao-Shan Gao. T2VSafetyBench: Evaluating the safety of text-to-video generative models. InAdvances in Neural Information Processing Systems, 2024

    2024

  23. [27]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of ...

    2021

  24. [28]

    Girshick, and Ali Farhadi

    Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 779–788, 2016. doi: 10.1109/CVPR.2016.91. URL https://doi.org/10.1109/CVPR.2016.91

    doi:10.1109/cvpr.2016.91 2016

  25. [29]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InThe Tenth International Conference on Learning Representations. OpenReview.net, 2022. URLhttps://openreview.net/forum?id=TIdIXIpzhoI

    2022

  26. [30]

    Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models

    Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522–22531,

  27. [31]

    Multi-task learning as multi-objective opti- mization

    Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective opti- mization. InAdvances in Neural Information Processing Systems, volume 31, pages 525–536, 2018. URL https://proceedings.neurips.cc/paper/2018/hash/ 432aca3a1e345e339f35a30c8f65edce-Abstract.html

    2018

  28. [32]

    Make-A-Video: Text-to-video generation without text-video data

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-A-Video: Text-to-video generation without text-video data. InThe Eleventh International Conference on Learning Representations. OpenReview.net, 2023. URLhttps://openreview.net/forum? id=...

    2023

  29. [34]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, T...

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.20314 2025

  30. [35]

    Towards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. URLhttp://arxiv.org/abs/1812.01717

    Pith/arXiv arXiv 2018

  31. [36]

    Erasing undesirable influence in diffusion models

    Jing Wu, Trung Le, Munawar Hayat, and Mehrtash Harandi. Erasing undesirable influence in diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28263–28273. Computer Vision Foundation / IEEE, 2025. doi: 10.1109/ CVPR52734.2025.02632. URL https://openaccess.thecvf.com/content/CVPR2025/html/ Wu_Erasing_U...

    arXiv 2025

  32. [37]

    Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient

    Yongliang Wu, Shiji Zhou, Mingzhuo Yang, Lianzhe Wang, Heng Chang, Wenbo Zhu, Xinting Hu, Xiao Zhou, and Xu Yang. Unlearning concepts in diffusion model via concept domain correction and concept preserving gradient. InProceedings of the AAAI Conference on Artificial Intelligence, pages 8496–8504. AAAI Press, 2025. doi: 10.1609/aaai.v39i8.32917. URLhttps: ...

    work page doi:10.1609/aaai.v39i8.32917 2025

  33. [38]

    VideoEraser: Concept erasure in text-to-video diffusion models

    Naen Xu, Jinghuai Zhang, Changjiang Li, Zhi Chen, Chunyi Zhou, Qingming Li, Tianyu Du, and Shouling Ji. VideoEraser: Concept erasure in text-to-video diffusion models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5954–5983. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025.emnlp-main....

    work page doi:10.18653/v1/2025.emnlp-main.304 2025

  34. [39]

    CogVideoX: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Yuxuan Zhang, Weihan Wang, Yean Cheng, Bin Xu, Xiaotao Gu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-video diffusion models with an expert transformer. InThe Thirteenth International Conference on Learning Represen...

    2025

  35. [40]

    T2VUnlearning: A concept erasing method for text-to-video diffusion models.arXiv preprint arXiv:2505.17550,

    Xiaoyu Ye, Songjie Cheng, Yongtao Wang, Yajiao Xiong, and Yishen Li. T2VUnlearning: A concept erasing method for text-to-video diffusion models.arXiv preprint arXiv:2505.17550,

    arXiv

  36. [41]

    SAFREE: Training- free and adaptive guard for safe text-to-image and video generation

    Jaehong Yoon, Shoubin Yu, Vaidehi Patil, Huaxiu Yao, and Mohit Bansal. SAFREE: Training- free and adaptive guard for safe text-to-image and video generation. InThe Thirteenth International Conference on Learning Representations. OpenReview.net, 2025. URLhttps: //openreview.net/forum?id=hgTFotBRKl

    2025

  37. [42]

    Forget-me- not: Learning to forget in text-to-image diffusion models

    Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me- not: Learning to forget in text-to-image diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1755–1764, 2024. doi: 10.1109/CVPRW63382.2024.00182. URL https://doi.org/10.1109/CVPRW63382.2024. 00182

    arXiv 2024

  38. [43]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018. doi: 10.1109/CVPR.2018.00068. URLhttps://doi.org/10.1109/CVPR.2018.00068

    arXiv 2018

  39. [44]

    To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images

    Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now. InProceedings of the European Conference on Computer Vision, 2024

    2024

  40. [45]

    VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    Dian Zheng, Ziqi Huang, Hongbo Liu, Kai Zou, Yinan He, Fan Zhang, Lulu Gu, Yuanhan Zhang, Jingwen He, Wei-Shi Zheng, Yu Qiao, and Ziwei Liu. VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025. doi: 10.48550/arXiv.2503.21755. URLhttps://doi.org/10.48550/arXiv.2503.21755

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.21755 2025

  41. [46]

    Open-Sora 2.0: Training a commercial-level video generation model in $200k.arXiv preprint arXiv:2503.09642,

    Zangwei Zheng, Xiangyu Peng, Yuxuan Lou, Chenhui Shen, Tom Young, Xinying Guo, Binluo Wang, Hang Xu, Hongxin Liu, Mingyan Jiang, Wenjun Li, Yuhui Wang, Anbang Ye, Gang Ren, Qianran Ma, Wanying Liang, Xiang Lian, Xiwen Wu, Yuting Zhong, Zhuangyan Li, Chaoyu Gong, Guojun Lei, Leijun Cheng, Limin Zhang, Minghao Li, Ruijie Zhang, Silan Hu, Shijie Huang, Xiaok...

    Pith/arXiv arXiv

  42. [51]

    URLhttps://arxiv.org/abs/2503.09642

    doi: 10.48550/arXiv.2503.09642. URLhttps://arxiv.org/abs/2503.09642. 16 Preprint Concept Unlearning for Text-to-Video Models A Detailed Related Work From image-domain concept erasure to video unlearning.A large portion of the generative unlearning literature was first developed for text-to-image diffusion models [6–8, 18, 25, 36, 37, 42]. These methods ty...

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.09642

  43. [2022]

    URL https://doi.org/10.48550/arXiv.2210.02303

    doi: 10.48550/arXiv.2210.02303. URL https://doi.org/10.48550/arXiv.2210.02303

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.02303

  44. [2023]

    URL https://doi.org/10.1109/CVPR52729

    doi: 10.1109/CVPR52729.2023.02157. URL https://doi.org/10.1109/CVPR52729. 2023.02157. 14 Preprint Concept Unlearning for Text-to-Video Models

    arXiv 2023

  45. [2024]

    URL https://doi.org/10.48550/arXiv.2407.16124

    doi: 10.48550/arXiv.2407.16124. URL https://doi.org/10.48550/arXiv.2407.16124. 13 Preprint Concept Unlearning for Text-to-Video Models

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.16124

  46. [2025]

    URLhttps://arxiv.org/abs/2505.17550

    doi: 10.48550/arXiv.2505.17550. URLhttps://arxiv.org/abs/2505.17550

    doi:10.48550/arxiv.2505.17550