acceptodds
Under review as a conference paper at ICLR 2027

GRAFT: Video Unlearning by Grafting Frames Temporally

Abstract

Text-to-video (T2V) diffusion models are trained on large-scale uncurated corpora and inherit the ability to render unsafe concepts such as explicit nudity, public figures, or specific objects. Existing unlearning methods for video adapt image-domain recipes: they steer the model's prediction away from a concept prompt, along a direction that is not tied to any video the model would produce. We observe that video offers a source of supervision that images lack: because frames in one generated video share subject, lighting, and camera, a concept-free version of a concept-bearing video can often be obtained by substituting frames rather than generating new ones. We propose GRAFT, which grafts concept-free latent frames over concept-bearing ones and fine-tunes toward the edited latent with the model's own objective. Because most concepts are never partially present, we further propose split-prompt generation, which conditions different temporal regions of one latent on a concept prompt and its safe counterpart, manufacturing partial-concept videos for any concept. We evaluate GRAFT on datasets spanning nudity, ImageNet objects, and celebrity identities, showing stronger erasure than prior work on CogVideoX and comparable erasure on HunyuanVideo, with aesthetic quality close to the base model.

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

Reviews

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

This paper proposes a method called GRAFT, which trains LoRA to achieve T2V concept erasure by replacing the latent frames containing the target concept with frames from the same generated video that are unrelated to the target concept, and then using the edited results to construct supervised data. For videos where it is difficult to naturally generate certain concepts, the authors propose a “split-prompt generation” method to construct training samples and validate the method’s effectiveness on tasks such as nudity removal, object removal, and character identity erasure.

Critical strengths

The method is explicitly motivated by video structure. Constructing erasure targets using safe donor frames from the same video is a distinctive supervised design; split-prompt generation, in turn, addresses the challenge of obtaining donor frames, and together, forms a relatively comprehensive methodological framework.

The results for Nudity Erasure on CogVideoX are outstanding. Table 1 compares publicly available T2VUnlearning checkpoints under the same evaluation process; GRAFT reduces the nude frame rate on Gen from 55.3% to 0.06%, compared to 13.0% for the baseline method. Results on prompt sets such as I2P and SafeSora also support its generalization capability.

Comprehensive experiments. The paper covers multiple concepts and models, and reports on sampler dependencies, degradation in technical quality, cases of failed identity erasure, and limitations of the metrics. These analyses help clarify the method’s scope of applicability and practical trade-offs.

Critical weaknesses

Lack of a stronger baseline comparison. The paper discusses VideoEraser in the “Related Work” section but does not include it in the experiments, nor does it explain why.

The evaluation of temporal quality is insufficient. I am curious whether GRAFT will cause the generated video to appear blurry or result in reduced dynamic range. Appendix E reports on VBench Subject Consistency and notes that this metric may be misleading due to blurriness. It is recommended to supplement this with metrics such as Motion Smoothness, Temporal Flickering, Dynamic Degree, and Background Consistency, and to combine these with human evaluation to distinguis true motion preservation.

The advantage of GRAFT on HunyuanVideo has diminished, but the reasons for this have not been analyzed. In Table 7, GRAFT’s exposure rates on Gen and Ring-A-Bell are 6.1% and 8.4%, respectively, both higher than T2VUnlearning’s 5.0% and 7.0%. Although GRAFT has better aesthetic quality, its erasure advantage has not been sustained. Whether the cause lies in stream-matching parameterization, architecture, training data, or hyperparameters?

Questions for authors

See weakness.

3 comments

Thank you for your helpful feedback. Every change to the paper is in blue; black text is unchanged from the submission.

The main change. Reviewer yurr pointed to the loss in technical quality, and 4Cje and you asked whether grafting degrades the video. We traced the lower measured quality to the training objective and fixed it (Section 3.5). The nudity row of Table 1 now comes from the revised objective, which the paper calls GRAFT; the submitted objective stays beside it as GRAFT-PF, so the trade can be read directly.

Nudity, recommended samplerOriginalT2VUnlearningSubmittedRevised
Gen nudity rate (%)55.313.00.060.06
Mean nudity rate, four sets (%)47.14.40.040.61
Mean DOVER-t, five sets0.0920.0650.0550.071
Mean DOVER-a, five sets0.9500.8280.9420.930

This brought a significant improvement in quality and motion, although both stay below the base model's (Appendix E.1). The revised objective covers nudity and the five identities on CogVideoX-5B, where it erases at least as well as the submitted target, and the ten object classes on CogVideoX-2B, where it erases about as well (restricted ESR-1 56.3 against 54.2, 1,000-way 94.3 against 96.2) and preserves more (restricted PSR-1 86.4 against 84.2, now above T2VUnlearning's 86.1). On identities it lowers the target's ID-Similarity by 99% on average (75% under the submitted target), changes the other identities' by 1% and raises their DOVER-t from 0.082 to 0.100 (base 0.105; Table 3). Fire and HunyuanVideo remain under the per-frame target, and the final version will report them under the revised objective.


VideoEraser. It was left out without explanation, which was our mistake. We ran the authors' released pipeline unchanged. On nudity (CogVideoX-5B) it does not lower the per-frame nudity rate: 59.4% on Gen and 58.3% on Ring-A-Bell, against 55.3% and 54.6% for the original model, in line with its authors' statement that it is more effective on concrete concepts. On chain saw (CogVideoX-2B) it erases: restricted ESR-1 60.7 against 68.3 for GRAFT and 11.5 for the original model, with similar preservation (restricted PSR-1 87.8 against 87.2). The results are in Appendix P.

Temporal quality. Appendix E.1 now reports the official VBench Motion Smoothness, Dynamic Degree, Background Consistency and Temporal Flickering (Table 7). GRAFT's Dynamic Degree on Gen falls from 45% to 10%, so its videos move less; its high smoothness, background consistency and flicker scores partly reflect that stillness. Where nothing is erased the loss is much smaller: on 40 held-out prompts of people with visible skin, Dynamic Degree is 58% against 78% at base and mean optical flow 1.87 against 2.32, 81% of the base model's (Appendix E). T2VUnlearning moves more (Dynamic Degree 64%) but scores below the base model on smoothness, background consistency and flicker, which points to unstable motion. The loss is concept-dependent: on identities GRAFT keeps motion, with optical flow on the erased identity's prompts within 15% of the original model's for four of five identities (Table 18), against -57% for GRAFT-PF.

As you suggest, we pair these metrics with a blind human evaluation (Appendix O). Raters who did not know which method produced a clip score GRAFT's motion as natural as T2VUnlearning's (2.53 and 2.38 on a 1–4 scale; the 95% interval rules out GRAFT being worse by more than 0.09). GRAFT's videos therefore move less without looking less natural. The clips of every method, including the base model, are on the anonymous page: https://graft-anonymous.github.io/graft/.

Why is HunyuanVideo weaker? Appendix F.1 now goes through the causes you list. Hyperparameters are not the cause: a factorial over the LoRA scope, learning rate, η\eta and timestep sampling shows a similar trade between erasure and damage in every run, and a rank-128 adapter on the same 40 blocks as T2VUnlearning's eraser behaves like our rank-8 one. We did not vary the training data, so it remains a possible cause; it does differ from the CogVideoX set, most notably in that a quarter of its rows ask for an edit of frames where the detector finds no nudity (Table 10). On parameterization and architecture, a fit probe shows the adapter making a large clip-wide change. We tested the structural difference between the parameterizations (an edit of size ησ\eta\sigma under rectified flow against ησ2\eta\sigma^2 under v-prediction): over the noise levels each model is trained on the two coefficients are nearly equal (0.89 and 0.90), so it does not explain the gap, and the averaged target of Eq. 5 scales the same way. We cannot separate the two further: no released video model shares HunyuanVideo's architecture under another parameterization. We had no GPU large enough to train HunyuanVideo during the rebuttal, so the revised objective has not been run on it.

Dear Reviewer,

we are thankful for your insightful feedback. As the end of the discussion period on Nov 18 is getting close, we would like to know if there are any additional points we should address. If there are none, we would be grateful if you could let us know whether you would consider raising your score.

Sincerely,

Authors

Thank you for the detailed response and additional experiments which substantially address the main concerns in my original review. Although reduced motion and the unverified applicability of the revised objective to HunyuanVideo remain limitations, the additional evidence strengthens the paper.

Reviewer_4CjeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRAFT suppresses target concepts in text-to-video diffusion models by constructing training targets from concept-free latent frames of the same generated video. Concept-bearing frames are replaced through interpolation or reflected donor segments, and a LoRA adapter is trained with a masked native velocity objective and a retention loss on concept-free anchors. Split-prompt generation creates videos in which different temporal regions contain or omit the concept, enabling the construction for concepts that appear only transiently.

Critical strengths

1、The supervision construction makes substantive use of video structure. Concept-free donor frames from the same video offer a plausible way to preserve scene context while defining an erasure direction. Split-prompt generation extends this idea beyond naturally transient concepts and is a useful technical component. The native velocity formulation also makes the construction compatible with the evaluated v-prediction and rectified-flow models.

2、The method achieves strong quantitative results on nudity erasure, and the evaluation methodology is commendable. The authors evaluate the publicly released T2VUnlearning checkpoint within their own pipeline using identical prompts, seeds, and detectors, rather than relying on numbers reported in separate papers. This supports a fairer comparison.

Critical weaknesses

1、This paper focuses on video generation, it does not present any video results. Without video examples, it is difficult to assess the claimed contributions and substantiate the reported findings.

2、Judging from the images presented in the paper, the generated videos appear to have limited visual quality. Moreover, showing only a single frame per video makes it difficult to assess the overall quality and temporal coherence. Equation 3 reverses and mirrors the donor segment, which actually gets a reversed motion that may not reflect natural motion in real videos. Equation 2 performs linear interpolation in the VAE latent space, which may introduce ghosting artifacts. Could these operations degrade the model’s original video generation capabilities?

Questions for authors

Could you provide more examples of face erasure? In particular, does erasing a face introduce distortions or other visual artifacts? Training a separate adapter and constructing a dedicated dataset for each target concept also seems cumbersome. Is there a way to unify this process across concepts?

6 comments

Thank you for your helpful feedback. Every change to the paper is in blue; black text is unchanged from the submission.

The main change. Reviewer yurr pointed to the loss in technical quality, and ecYz and you asked whether grafting degrades the video. We made an effort to trace why the measured quality is lower and found the cause in the training objective, which we subsequently fixed (Section 3.5). The nudity row of Table 1 now comes from the revised objective, which the paper calls GRAFT; the submitted objective stays beside it as GRAFT-PF, so the trade can be read directly.

Nudity, recommended samplerOriginalT2VUnlearningSubmittedRevised
Gen nudity rate (%)55.313.00.060.06
Mean nudity rate, four sets (%)47.14.40.040.61
Mean DOVER-t, five sets0.0920.0650.0550.071
Mean DOVER-a, five sets0.9500.8280.9420.930

This brought a significant improvement in quality and motion, although both stay below the base model's (Appendix E.1). The revised objective covers nudity and the five identities on CogVideoX-5B, where it erases at least as well as the submitted target, and the ten object classes on CogVideoX-2B, where it erases about as well (restricted ESR-1 56.3 against 54.2, 1,000-way 94.3 against 96.2) and preserves more (restricted PSR-1 86.4 against 84.2, now above T2VUnlearning's 86.1). On identities it lowers the target's ID-Similarity by 99% on average (75% under the submitted target), changes the other identities' by 1% and raises their DOVER-t from 0.082 to 0.100 (base 0.105; Table 3). Fire and HunyuanVideo remain under the per-frame target, and the final version will report them under the revised objective.


No video results. Videos for every method and prompt set are on an anonymous page, with nudity masked: https://graft-anonymous.github.io/graft/. In the paper, Figure 5 now shows five frames across each clip rather than one.

Could Equations 2 and 3 damage motion? We found your point insightful: as you said, Eq. 3 mirrors the donor segment and so asks the model for reversed motion. We therefore fixed the objective, and the adapter no longer regresses the donor frames one by one: its target is the base model's own prediction plus the edit averaged over the concept frames (Eq. 5), so the order of the donor frames no longer enters the target and the motion comes from the base model. The edit is still taken from the same video, as the mean difference between the donor frames and the concept frames they replace at each spatial position. Averaging alone was not enough; it took the edit-only form as well (Table 21), and together they recover technical quality and motion over the per-frame target. Eq. 2 does not apply to the split-prompt data, because split-prompt concept regions always touch the clip's edge; it was used only for fire. We have no ghosting metric, so we make no claim; the video page lets you judge this directly.

More face examples and distortions. More face examples are on our video page. They are full videos, which give a better feel for the results than single frames. On distortions, Appendix J shows the dataset-dependent failure: a varied dataset blurs or smears the face, while the reported dataset replaces it with an intact face. Under the revised objective the erased identity is recognized in none of its 30 videos, while a face stays present, and the other identities' DOVER-t returns to 0.100 (base 0.105). We also ran a blind human study on your question. Raters, not aware of the method or prompt a video comes from, rated GRAFT's identity-erased videos as having the same level of distortion as the original model's (2.52 against 2.57 on a 1–4 scale; the 95% interval rules out a difference above 0.24; Appendix O).

Adapting the method to many concepts. The pipeline is the same for every concept, and a new concept needs only a list of (A,B,C)(A, B, C) prompt triples. Split-prompt generation fixes the mask in advance, so no concept detector enters training, and the training settings are shared across the classes of a concept type (Appendix A); building a dataset is a generation job on the base model followed by screening. Since each training row carries its own prompt, mask and edit, the datasets of several concepts can also be merged and one adapter trained on their union, with no change to the method. We tested this on two objects: one adapter trained on the union of the tench and English-springer rows (400 steps, otherwise the recipe of Section 4.3) erases both about as well as two separate adapters: restricted ESR-1 is 72.8 and 89.1 against 78.8 and 94.7, within the run-to-run spread of about 7 points, and 1,000-way ESR-1 is 100.0 for both, while the other eight classes are preserved better (restricted PSR-1 91.4 against 86.8 and 83.9; Appendix L). Scaling this to many concepts at once is future work.

Dear Reviewer,

Thanks again for your feedback. Since the Nov 18 deadline is approaching, we would like to ask if you need any additional clarifications from us. In case our comment has already addressed your concerns, please let us know whether you would consider updating your assesment.

Sincerely,

Authors

1、The videos you provided appear to be of relatively low quality. In many cases, the visual content changes substantially, and the characters’ movements also differ from the originals.

2、 If the goal is simply to remove unsafe concepts, wouldn’t screening the input text be simpler and more direct, especially if your method cannot preserve the original motion? For example, you can just replace unsafe concepts text with a new text. Moreover, building a separate training dataset for each unsafe concept is laborious and difficult to scale in practice.

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

The paper proposes GRAFT, a concept-erasure method for text-to-video diffusion models. It constructs training targets by replacing concept-bearing latent frames with concept-free frames from the same generated video, then fine-tunes a LoRA adapter toward these targets. Split prompt generation produces suitable training videos when concepts do not naturally appear only temporarily. Experiments cover nudity, objects, and identities on CogVideoX, with additional nudity results on HunyuanVideo. GRAFT achieves strong erasure, although technical quality scores decline.

Critical strengths
  1. Using frames from the same video to construct concept-free supervision is a useful contribution specific to video generation. Split prompt generation makes this approach applicable beyond concepts that naturally appear and disappear. The method is reasonably simple and supports both diffusion and flow-matching models.

  2. The CogVideoX nudity results are convincing. On Gen, GRAFT reduces the frame-level nudity rate from 55.3% to 0.06%, compared with 13.0% for the released T2VUnlearning checkpoint evaluated in the same pipeline. Results across additional prompt sets, objects, and identities suggest that the benefit extends beyond one benchmark.

Critical weaknesses
  1. The loss in technical quality, measured by DOVER-t, is substantial and affects some non-target outputs. Under the recommended sampler, the mean DOVER-t score across the nudity sets falls from 0.092 to 0.055. For identity erasure, the mean DOVER-t score on the other identities falls from 0.105 to 0.082 despite their identity similarity being preserved. The inference gate improves DOVER-t scores, but restores considerable nudity on Ring-A-Bell. These results weaken the broader claim that non-target content remains close to the base model, and the 15 unrelated prompts provide only limited reassurance.

  2. The main grafting ablation does not isolate the proposed contribution. Appendix K changes both the input videos and the supervision mask when replacing grafted targets with independently generated targets: GRAFT uses split-prompt inputs and a temporal mask, whereas the alternatives use ordinary A generations and supervise every frame. Their collapse could therefore reflect these changes rather than target construction alone. Also, the experiment only covers chain saw, which gives limited insight.

Questions for authors
  1. Can you provide a comparison that holds the input videos, temporal mask, loss normalization, and unlearning strength fixed while changing only the target construction? This would help establish whether the advantage comes specifically from using donors from the same video.

  2. How well does the nudity adapter preserve benign human content involving exposed skin, such as swimming, sports, or portraits? This matters because skin-containing retention anchors are deliberately excluded, and DOVER-t scores on the paired safe prompts also decrease.

3 comments

Thank you for your helpful feedback. Every change to the paper is in blue; black text is unchanged from the submission.

The main change. You pointed to the loss in technical quality, and 4Cje and ecYz asked whether grafting degrades the video. We made an effort to trace why the measured quality is lower and found the cause in the training objective, which we subsequently fixed (Section 3.5). The nudity row of Table 1 now comes from the revised objective, which the paper calls GRAFT; the submitted objective stays beside it as GRAFT-PF, so the trade can be read directly.

Nudity, recommended samplerOriginalT2VUnlearningSubmittedRevised
Gen nudity rate (%)55.313.00.060.06
Mean nudity rate, four sets (%)47.14.40.040.61
Mean DOVER-t, five sets0.0920.0650.0550.071
Mean DOVER-a, five sets0.9500.8280.9420.930

This brought a significant improvement in quality and motion, although both stay below the base model's (Appendix E.1). The revised objective covers nudity and the five identities on CogVideoX-5B, where it erases at least as well as the submitted target, and the ten object classes on CogVideoX-2B, where it erases about as well (restricted ESR-1 56.3 against 54.2, 1,000-way 94.3 against 96.2) and preserves more (restricted PSR-1 86.4 against 84.2, now above T2VUnlearning's 86.1). On identities it lowers the target's ID-Similarity by 99% on average (75% under the submitted target), changes the other identities' by 1% and raises their DOVER-t from 0.082 to 0.100 (base 0.105; Table 3). Fire and HunyuanVideo remain under the per-frame target, and the final version will report them under the revised objective.


DOVER-t loss, and non-target content. Mean DOVER-t on the nudity sets rises from 0.055 to 0.071 (base 0.092, T2VUnlearning 0.065), and the revised checkpoint no longer needs the inference gate you questioned. DOVER-t on the paired safe prompts rises from 0.053 to 0.070. You are right that 15 unrelated prompts give limited reassurance, so we added 40 held-out prompts of people in everyday scenes with visible skin (swimming, beach, sports, portraits, dance). On all 55 unrelated prompts DOVER-t reads 0.092 against 0.097 at base (T2VUnlearning 0.091) and DOVER-a 0.961 against 0.970 (0.949); on the 40 new prompts GRAFT's mean optical flow is 81% of the base model's (1.87 against 2.32; Appendix E). We have softened the claim in the Discussion accordingly.

Isolating the target construction. You are right that the Appendix K ablation changed the inputs and the mask as well as the target. We have now run the comparison you asked for, on nudity under the revised objective. Everything is held fixed (input videos, temporal mask, loss normalization, η\eta, training steps) and only the source of the replacement frames changes: from the same video to a second video that shares its seed and early denoising steps but is generated with the concept-free prompt. With frames from the second video, 16.8% of Gen frames stay nude, against 0.06% with frames from the same video (T2VUnlearning 13.0%), at slightly higher technical quality (DOVER-t 0.081 against 0.071). So on nudity the advantage comes from taking the replacement frames from the same video.

We also added a second comparison, the forked pair: the two versions of a clip share their early denoising steps, and the target is the whole concept-free version rather than a graft. On nudity GRAFT-PF stays clearly ahead (nudity rate 0.00–0.09% against 0.15–2.00% on four sets for the forked pair's best run; DOVER-a 0.93–0.97 against 0.36–0.50), and by eye the forked pair smears the subject. On the ten ImageNet classes it ties GRAFT-PF (PSR-1 44.3 for both, restricted ESR-1 53.9 against 54.2), so for objects we do not claim an advantage from grafting (Appendix K).

Benign exposed skin. We understand the concern, and a method that separates nudity from benign exposed skin, such as swimwear, would be useful future work. Our nudity protocol follows T2VUnlearning's, whose detector counts exposed belly, armpits and feet as nudity, so erasing them is part of the task as the benchmark defines it; setting the boundary between nudity and benign skin is arbitrary, so we followed the existing procedures. Preserving benign exposed skin was therefore not a goal of this design, which is also why such anchors are excluded from retention. The adapter does preserve people: on the 79 clothed safe rewrites subject and scene are kept, at DOVER-t 0.070 against 0.087 at base (T2VUnlearning 0.075), and the video page (https://graft-anonymous.github.io/graft/) shows these rewrites for every method, so the rendering of people can be judged directly.

Dear Reviewer,

we are grateful for your comment. Since the Nov 18 deadline is approaching, we would like to find out whether you need any additional clarifications from us. If not, please indicate whether you would consider updating your assessment.

Sincerely,

Authors

I thank the authors for their comprehensive rebuttal. The training objective fix, in particular, gives more stable results, which lowers my concern about the technical quality loss, even if loss is still present. The new ablations also provide more clarity about the main contribution of the GRAFT method. My score was already positive, and therefore remains so.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.