FADE: Frame-Aware Diffusion-Transformer-based Multi-Concept Erasure for Video Unlearning

Yuchen Li; Kaiyuan Deng; Chaoran Feng; Zhenyu Tang; Li Yuan

cs.CVOct 2, 2026

FADE: Frame-Aware Diffusion-Transformer-based Multi-Concept Erasure for Video Unlearning

Authors: Yuchen Li, Kaiyuan Deng, Chaoran Feng, Zhenyu Tang, Li Yuan

Organizations: Peking University · University of Arizona

Abstract

Text-to-video (T2V) diffusion models can reproduce copyrighted, violent, or explicit content, which motivates concept erasure: removing designated concepts from a pretrained model while preserving its behavior on everything else. Existing T2V erasure methods leave two problems open. Their frame-agnostic suppression can leave isolated frames in which an erased concept resurfaces, a frame-reactivation gap that clip-level averages obscure; and they are usually evaluated with one target concept or category at a time. We propose Frame-Aware Diffusion Erasure (FADE), a multi-concept video unlearning framework. FADE first applies a joint closed-form key/value edit that suppresses all target concepts, then trains per-concept frame-aware low-rank adapters whose strength is gated by the frame index and the denoising timestep to remove residual per-frame leakage. Each adapter is trained with the other targets' prompts as hard negatives, which keeps the concept-specific components of different adapters well separated, and a similarity-based soft router combines the adapters according to the prompt. With 16 concepts (objects, artistic styles, and nudity) erased from a single Wan2.1-T2V-1.3B backbone, FADE reduces the residual accuracy on the object benchmark to 4.9%, against 15.5% for the strongest of eight baselines, while keeping the VBench average within 0.9% of the unedited model. The ranking is unchanged under a VLM judge and a blinded human study, and the advantage over the strongest baseline carries over to prompts that combine several erased concepts, to 30 simultaneously erased celebrity identities, and to Wan2.1-T2V-14B, CogVideoX-2B, and HunyuanVideo-1.5.

Figures & tables

Figure 1: Frame reactivation and multi-concept erasure. Prior T2V erasure methods apply the same suppression to every frame, and an erased concept can resurface mid-clip (red boxes; top: Refusal Vector on parachute , middle: VideoEraser on English springer ). FADE erases several concepts from a single backbone, and in these examples they stay suppressed in every frame (bottom).

Figure 2: FADE overview. (a) Concept Redirect: motion-paraphrased concept embeddings enter a joint closed-form K/V solve, and TPD blends the edited and original projections over timesteps. (b) FAE training for one concept (here parachute ): the unsafe prompts of every other erased concept join the neutral set of $\mathcal{L}_{\mathrm{motion}}$ as hard negatives; a frozen CR teacher supplies the targets. (c) At inference, FrameGate modulates the frame-aware branch of each FA-LoRA and the C-MoE router scales each expert by its prompt gate $g_{c}$ .

Figure 3: FA-LoRA attached to the first FFN linear of a DiT block. The static branch $B_{0}D_{0}$ and the frame-aware branch $B_{1}(\phi_{c}\odot D_{1}\,\cdot)$ are added to the frozen projection; FrameGate $\phi_{c}$ is shared across the blocks of one concept.

	Per-class $\mathrm{mean\text{-}frame\text{-}acc}$ (%) $\downarrow$
Method	Cassette player	Chain saw	Church	English springer	French horn	Garbage truck	Gas pump	Golf ball	Parachute	Avg acc $\downarrow$	FVD $\downarrow$
Wan2.1 (unedited)	56.6	72.2	78.4	55.4	96.9	78.2	64.9	99.1	98.3	77.8	147.3
SAFREE	43.2	40.3	45.2	36.5	69.4	44.6	36.5	77.8	80.1	52.6	213.1
T2VUnlearning	49.5	65.6	68.3	39.7	79.5	57.3	53.6	89.7	88.4	65.7	215.8
Refusal	38.7	54.9	58.2	36.3	72.3	60.5	59.4	81.7	78.6	60.1	184.9
VideoEraser	12.4	19.3	17.8	11.5	18.6	11.2	18.5	15.9	14.2	15.5	203.4

Table 1: Object erasure on Imagenette , all classes erased from one Wan2.1-1.3B model (ResNet-50 judge; FVD on the neutral MSR-VTT pool). Tench is excluded because Wan2.1 cannot generate it.

Figure 4: Qualitative results on six concepts from the object, style, and nudity categories, all erased by one 16-concept Wan2.1 model. Three frames per clip; struck-through words mark erased concepts.

Table 2: Additional baselines and judges on the benchmark of Table 1 . Human: blinded study, mean 0–4 rating rescaled to 0–100.

Table 3: (a) VBench fidelity , neutral MSR-VTT pool. (b) Style erasure , per-frame CLIP similarity.

Table 4: Multi-concept erasure. (a) Prompts mentioning two erased classes, scored per target. (b) FADE erasing nested subsets of

N

Imagenette classes.

Figure 5: Compositional erasure. Prompts that mention two erased concepts (struck through); the 16-concept model removes both while preserving the layout, lighting, and motion of the scene.

	Avg $\mathrm{mean\text{-}frame\text{-}acc}$ (%) $\downarrow$			FADE vs. unedited		Nudity (%) $\downarrow$
Backbone	Unedited	VideoEraser	FADE	Reduction	FVD incr.	Unedited	VideoEraser	FADE
Wan2.1-T2V-1.3B	84.8	16.2	5.1	94.0%	+7.8%	84	42	19
Wan2.1-T2V-14B	90.3	21.3	7.2	92.1%	+13.6%	86	39	15
CogVideoX-2B	81.5	18.3	6.3	92.3%	+8.4%	79	49	22
HunyuanVideo-1.5-480P	72.6	15.3	9.2	87.3%	+17.2%	–	–	–

Table 5: FADE on four backbones : five Imagenette classes shared by all backbones and NudeNet-flagged I2P frames. Ratios are computed before rounding.

Table 11

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Imagenette object	$\tau_{c}$	$T_{c}$
cassette player	0.320	0.050
chain saw	0.347	0.030
church	0.336	0.020
English springer	0.455	0.033
french horn	0.279	0.045
garbage truck	0.413	0.022

Appendix

Table 8: Per-concept C-MoE routing thresholds. Calibrated on a held-out unsafe/neutral split with cross-concept hard negatives, so that

\tau_{c}

separates in-class from neutral and out-of-class prompts. The fallback row

\tau_{c}{=}1.000

marks the regime where the median gap collapses (see text).

Table 9: (a) Unsafe content on I2P , frames flagged by NudeNet (nudity) or Q16 (violence, hate). (b) 30 erased celebrities : GIPHY detector accuracy on erased (

\mathrm{Acc}_{e}

) and retained (

\mathrm{Acc}_{s}

) identities.

Class	CogVideoX-2B	Refusal	VideoEraser	FADE
Chain Saw	68.5	51.3	21.7	6.8
Church	75.1	55.6	19.4	5.5
French Horn	93.7	68.9	20.3	6.7
Garbage Truck	74.6	57.2	13.8	5.3
Parachute	95.8	74.1	16.5	7.2
Avg $\mathrm{mean\text{-}frame\text{-}acc}\!\downarrow$	81.5	61.4	18.3	6.3

Appendix

Table 10: Object erasure on CogVideoX-2B. Per-class

\mathrm{mean\text{-}frame\text{-}acc}

(%)

\downarrow

over 5 prompts

\times

5 seeds per class, their average, and FVD on the neutral MSR-VTT pool.

Class	Wan2.1-14B	Refusal	VideoEraser	FADE
Chain Saw	82.5	63.7	24.6	7.8
Church	86.3	67.4	22.1	6.2
French Horn	98.6	79.5	23.8	7.5
Garbage Truck	85.1	68.3	16.4	6.0
Parachute	99.1	84.2	19.7	8.3
Avg $\mathrm{mean\text{-}frame\text{-}acc}\!\downarrow$	90.3	72.6	21.3	7.2

Appendix

Table 11: Object erasure on Wan2.1-T2V-14B. Same protocol as Table 10 .

Panel of Figure 6 — Figure 6: Nudity erasure on I2P across three backbones. Frames flagged by NudeNet (%).

FrameGate variant	Params	$\mathrm{mean\text{-}frame\text{-}acc}\!\downarrow$	FVD $\downarrow$
2-layer MLP, width 32 (default)	290	4.9	158.8
4-layer MLP, width 128	34,178	6.3	196.7
Temporal 1-D convolution over frames	802	4.5	254.0
Frame index only (no $\bar{t}$ )	226	7.5	220.3
Timestep only (no $\bar{f}$ )	226	6.1	204.2

Appendix

Table 12: FrameGate variants (ten Imagenette experts, Wan2.1-1.3B). Only the gate changes.

Stage	Cost	Dependence on $N$
CR closed-form edit	63.9 s ( $N{=}1$ ), 65.9 s ( $N{=}5$ ), 67.0 s ( $N{=}12$ )	Nearly constant: one joint solve with one matrix inverse per projection; runtime is dominated by loading and saving the model
FAE training	3 h 50 min–3 h 58 min per concept (all 16 concepts); 500 AdamW iterations, about 28 s per iteration	Linear; for a fixed concept set, runs are independent and parallelizable
C-MoE calibration	Minutes per concept set; forward passes only (encoder scores for about 530 prompts per concept plus the other concepts’ prompts, then quantiles)	Cached encoder work grows linearly with the $500{+}30N$ prompts; similarity scoring grows as $O(N(500{+}30N))$ , negligible at $N{=}16$
Storage	0.56M trainable parameters (1.15 MB) per concept; 18.4 MB for 16 experts	Linear, negligible next to the 2.6 GB backbone
Inference	7.29 s per step with no active expert (same as the unedited model); 7.62 s per step with two active experts (+4.5%)	Per-step cost is independent of $N$ for a fixed number of active experts (about +2% each); routing computes $N$ similarities once per prompt

Appendix

Table 13: Cost of FADE by stage and its dependence on the number of concepts

N

Figure 7: Additional qualitative comparison (part 1). Struck-through words mark erased concepts.

Figure 8: Additional qualitative comparison (part 2). Continuation of Fig. 7 . The Studio Ghibli example uses an additional style expert outside the 16-concept benchmark.

Diffusion Model Unlearning

Diffusion model unlearning removes specified concepts or training examples from generative models while preserving unrelated generation.

See all 72 papers

Motion Concept Unlearning in Video Diffusion Models
Sep 29, 2026Ping Liu, Chi ZhangVideo Diffusion Models Diffusion Transformer
Computer Science and Engineering University of Nevada, Reno Reno, NV, USA · Department of Mathematics National University of Singapore Singapore
PROBE: Diagnosing Residual Concept Capacity in Erased Text-to-Video Diffusion Models
Mar 23, 2026Yiwei Xie, Zheng Zhang, Ping LiuDiffusion Model Unlearning Video Generation Evaluation
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China · Department of Computer Science and Engineering, University of Nevada, Reno, NV, USA
EraseAnything++: Enabling Concept Erasure in Rectified Flow Transformers Leveraging Multi-Object Optimization
Mar 1, 2026Zhaoxin Fan, Nanxiang Jiang, Daiheng Gao +2Concept Erasure Diffusion Model Unlearning
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University. · University of Science and Technology of China.

Sep 29, 2026J/K move · Enter open · S save

Motion Concept Unlearning in Video Diffusion Models

Ping Liu, Chi Zhang

Computer Science and Engineering University of Nevada, Reno Reno, NV, USA · Department of Mathematics National University of Singapore Singapore

Video Diffusion Models Diffusion Transformer Attention Control in Diffusion Models Diffusion Model Unlearning Machine Unlearning

Open paper AskPDF arXiv

Text-to-video (T2V) diffusion models can generate realistic depictions of actions such as kicking, stabbing, and shooting, raising safety concerns that motivate targeted concept erasure. Although concept erasure has been extensively studied for static concepts in text-to-image and T2V models, erasing motion concepts remains largely unexplored. We present a systematic study of motion concept erasure in video Diffusion Transformers (DiTs). Through causal interventions, we show that text-conditioning attention carries concept-specific motion information and supports selective intervention, whereas perturbing temporal positional encoding suppresses both target and non-target dynamics. We further find that directly adapting ESD, a representative weight-level image erasure method, to a video DiT yields modest and uneven motion suppression: reducing its erasure training loss does not by itself remove the concept signal from the difference between the conditional and unconditional predictions, which classifier-free guidance (CFG) then scales at every denoising step. From these findings, we derive three requirements for motion concept erasure: concept specificity, spatial selectivity, and temporal naturalness. Each determines one component of MUTE (Motion concept Unlearning in Text-to-video gEneration): at each denoising step, MUTE extracts a concept direction through token neutralization, derives a spatial gate from the direction's intrinsic structure, and subtracts the resulting correction from the velocity output before CFG is applied. MUTE is training-free and requires no weight modification. Experiments on 20 motion concepts show that MUTE outperforms representative prompt-level, weight-level, and inference-time baselines on Wan2.1-T2V, and the same formulation transfers to CogVideoX, supporting its applicability across distinct T2V attention architectures.

Abstract

Diffusion Model Unlearning

Explore similar work

Motion Concept Unlearning in Video Diffusion Models

PROBE: Diagnosing Residual Concept Capacity in Erased Text-to-Video Diffusion Models

EraseAnything++: Enabling Concept Erasure in Rectified Flow Transformers Leveraging Multi-Object Optimization