MUSE: Mask Union for Stable Erasure

• The paper presents MUSE, a temporal mask union strategy that preserves all object positions during downsampling, thereby mitigating localization failures.
• It enhances video inpainting by computing an element-wise logical OR over grouped masks, ensuring transient object regions are not missed.
• Empirical results show that MUSE improves metrics like PSNR, SSIM, and LPIPS, especially under conditions of temporal sparsity and abrupt motion.

Searching arXiv for papers on “Mask Union for Stable Erasure” and closely related mask-based erasure methods. Mask Union for Stable Erasure (MUSE) is a component of the Stable Video Object Removal (SVOR) framework that addresses a specific instability in diffusion-based video object removal: localization failure caused by temporal mask compression under abrupt motion and imperfect mask availability. In SVOR, MUSE is defined as a windowed union strategy applied during temporal mask downsampling. Instead of selecting a single frame mask to represent a temporal window, it retains the union of all mask locations observed in the window, thereby preserving short-lived object positions that would otherwise be dropped. The method is deterministic, introduces no extra trainable parameters, and is designed to reduce incomplete removal, under-erasure, ghosting, smearing, and frame-to-frame flicker when masks are sparse, inaccurate, or partially missing (Hu et al., 10 Mar 2026).

1. Problem setting and motivation

Video object removal takes a video together with a mask indicating the object to erase, and seeks to remove that object while reconstructing a plausible background that is spatially coherent and temporally stable. SVOR emphasizes that this setting becomes substantially harder under real-world imperfections such as shadows, abrupt motion, and defective masks. In deployment, masks may be temporally sparse, spatially inaccurate, or partially missing because of segmentation failures, annotation sparsity, occlusion, motion blur, or fast object motion (Hu et al., 10 Mar 2026).

The specific failure targeted by MUSE arises because diffusion-based video inpainting and removal pipelines often do not keep masks at full temporal resolution throughout the model. Instead, masks are temporally compressed to align with the latent video representation. Under abrupt motion, an object may occupy very different positions across neighboring frames within one compression window. If the pipeline keeps only one sampled frame from that window, some true object locations disappear from the compressed mask. The downstream model is then never told to erase those missed regions. SVOR characterizes this as a structural misalignment problem caused by temporal mask compression, rather than merely insufficient model robustness (Hu et al., 10 Mar 2026).

This diagnosis is central to MUSE. The method is not presented as a general-purpose mask heuristic, nor as morphology, dilation, or learned propagation. Its scope is narrower and more specific: preserve localization coverage during temporal downsampling so that every observed target location within a compression window survives into the latent-time mask sequence that conditions the video model.

2. Formal mechanism

Let the input mask sequence be

$\mathbf{M}=\{M_1,M_2,\dots,M_T\},$

where each $M_t$ is a spatial binary mask for frame $t$. Let the temporal compression ratio be $r$, which is 4 by default in the SVOR implementation. Standard temporal downsampling would typically keep one mask every $r$ frames. MUSE instead uses first-frame anchoring plus grouped temporal union (Hu et al., 10 Mar 2026).

The first compressed mask is preserved directly: $M_1^{\downarrow}=M_1.$ For later compressed indices $j\ge 2$, the downsampled mask is formed by element-wise union over a temporal group: $M_j^{\downarrow}=\bigvee_{t\in\mathcal{G}_j} M_t,$ where $\bigvee$ denotes element-wise logical OR over binary masks. Equivalently, for spatial coordinates $(h,w)$,

$M_t$0

The paper’s implementation description is procedural. Starting from

$M_t$1

SVOR: preserves the first frame mask exactly, partitions the remaining frames into non-overlapping temporal groups of size $M_t$2, computes an element-wise union for each group, and feeds the resulting compressed sequence

$M_t$3

into the latent-conditioning pipeline (Hu et al., 10 Mar 2026).

The key mechanism is therefore simple: MUSE answers the question “which pixels were ever masked anywhere in this window?” rather than “which one mask should represent this window?” SVOR describes this as a coverage-preserving temporal mask downsampling rule.

3. Position inside SVOR

SVOR consists of three main designs: Mask Union for Stable Erasure (MUSE), Denoising-Aware Segmentation (DA-Seg), and Curriculum Two-Stage Training. Their roles are complementary rather than interchangeable (Hu et al., 10 Mar 2026).

Component	Function	Scope
MUSE	preserves all target regions within each temporal window	temporal mask downsampling
DA-Seg	provides an internal diffusion-aware localization prior	decoupled side branch
Curriculum Two-Stage Training	learns realistic background and temporal priors, then refines on synthetic pairs	training strategy

In SVOR, the model is built on Wan2.1-VACE-1.3B with a lightweight context branch that injects input frames and masks into the DiT backbone. MUSE modifies the temporally compressed mask fed through that conditioning route. It therefore changes the effective erasure guidance region seen by the model, but it does not alter the denoising architecture itself, add an extra loss, or introduce a trainable block (Hu et al., 10 Mar 2026).

SVOR states that MUSE is applied during Stage II training and inference. This matters because the model is trained under the same temporally unioned conditioning that it later encounters at test time. The supplementary material also describes a training-free, plug-and-play inference-time enhancement for prior methods: masks are temporally grouped and unioned, and the resulting union mask is repeated to recover the original frame count. That workaround is explicitly presented as an approximation for pre-existing models, whereas in SVOR proper the union is integrated at the mask downsampling stage aligned with latent compression (Hu et al., 10 Mar 2026).

4. Empirical evidence

The empirical case for MUSE is presented as both qualitative and quantitative. Qualitatively, SVOR’s “stable_abrupt” comparison shows that without MUSE, removal fails on abrupt-motion frames, whereas with MUSE integrated the model achieves stable removal. The same comparison also indicates that inference-only MUSE improves results, but the strongest behavior comes when MUSE is part of the training-and-inference pipeline (Hu et al., 10 Mar 2026).

On ROSE Bench, the ablation starting from the Stage II base model trained only on paired ROSE data reports the following changes after adding MUSE:

Metric	Stage II base	+ MUSE
PSNR	27.33	27.87
SSIM	0.918	0.921
LPIPS	0.066	0.061
ReMOVE	0.901	0.902

The absolute ReMOVE improvement is small, and SVOR explicitly notes that abrupt frame displacement is relatively rare in standard test sets, so average benchmark metrics understate MUSE’s value. Even so, MUSE improves all reported image-quality metrics and slightly improves removal quality while adding no extra parameters (Hu et al., 10 Mar 2026).

The most targeted quantitative evidence comes from controlled temporal sparsification. A frame-skipping factor $M_t$4 is used to simulate increasing temporal discontinuity. As $M_t$5 increases, MUSE becomes more beneficial. For example, on gen-omni, ReMOVE changes from 0.8998 to 0.9003 at $M_t$6, 0.9011 to 0.9020 at $M_t$7, and 0.9014 to 0.9027 at $M_t$8. On minimax, the change is 0.8960 to 0.8955 at $M_t$9, 0.8968 to 0.8987 at $t$0, and 0.9001 to 0.9032 at $t$1. On ROSE, the change is 0.9081 to 0.9080 at $t$2, 0.9060 to 0.9068 at $t$3, and 0.9016 to 0.9024 at $t$4 (Hu et al., 10 Mar 2026).

SVOR draws two conclusions from these results. First, MUSE has negligible negative effect when abrupt motion is absent. Second, its benefit increases as temporal sparsity or abruptness becomes stronger. This is the expected signature for a mechanism whose purpose is to prevent mask collapse during temporal compression.

5. Interpretation, trade-offs, and limitations

The intuition given for MUSE is direct: the compressed latent step should know all places where the target might exist during the corresponding temporal window. If only one frame’s mask is kept, transient positions are lost. If the union of all positions visited within that window is used instead, the model cannot miss a location merely because it did not appear in the sampled frame. SVOR describes this as converting a brittle point estimate of object location into a conservative temporal envelope over the window (Hu et al., 10 Mar 2026).

That conservatism is both the strength and the main trade-off of MUSE. The benefit is fewer missed removals and less flicker. The risk is that the effective erase region can become larger than strictly necessary. SVOR explicitly notes that under extreme sparsity, such as only a single-frame mask, the system may cause false/over-erasure. This is highly relevant to MUSE because temporal union prefers preserving all potentially target-associated locations. If the object occupies different positions across time, or if masks are noisy, the union can over-expand the region to be inpainted (Hu et al., 10 Mar 2026).

The paper does not report significant computational or memory overhead for MUSE, and the operation itself is cheap relative to diffusion inference. The main design sensitivity is the temporal compression ratio, since the union is taken over each compression window. SVOR uses the model’s native compression ratio, 4 by default, rather than introducing a separate tunable window size (Hu et al., 10 Mar 2026).

A common misconception is to read MUSE as a learned temporal aggregation module. SVOR does not describe it that way. MUSE is a deterministic mask-space preprocessing/downsampling operation. It is not a latent-feature union, not an attention-feature union, and not a learned segmentation mechanism.

6. Relation to adjacent “stable erasure” research

The term “MUSE” is ambiguous in current arXiv literature. In Fast Unlearning at Scale via Margin Self-Correction, MUSE News and MUSE Books refer to a language-model unlearning benchmark, not to Mask Union for Stable Erasure. That paper is therefore relevant only as a nomenclature clarification: its MUSE is Machine Unlearning Six-Way Evaluation for LLMs, not the temporal mask-union mechanism introduced in SVOR (Gennaro et al., 1 Jun 2026).

Among adjacent mask-based erasure methods, PECKER is directly relevant only in an analogous sense. It studies diffusion-model machine unlearning and uses a binary, thresholded, gradient-based saliency mask over parameters so that forget batches update only masked parameters. The paper explicitly states that it does not present the mask as a union of multiple masks across concepts, timesteps, or tasks. PECKER therefore supports the broader idea that masking can stabilize selective erasure, but it does not implement MUSE-style temporal union (Ma et al., 7 Apr 2026).

Mosaic is closer in spirit to multi-mask composition. It constructs concept-specific masks from discrepancies between a base vector field and concept-erased vector fields, then blends concept-erased fields locally during sampling. However, its composition is not a simple binary OR union; it uses discrepancy normalization, relative discrepancy ratios, thresholded assignment masks, and a base fallback region. Mosaic is thus conceptually close to a dynamic, conflict-aware mask-composition framework, but it is not the same mechanism as MUSE’s grouped temporal logical OR during downsampling (Ko et al., 25 May 2026).

Other concept-erasure methods reinforce the same distinction. SPACE induces sparse support in cross-attention through iterative $t$5-regularized updates, yielding an implicit support mask rather than an explicit union mask (Novello et al., 11 May 2026). Z-Erase uses a token-stream gate

$t$6

to confine LoRA adaptation to text-token routes in single-stream diffusion transformers; this is a structured update mask, not a union over temporal masks (Jiang et al., 26 Mar 2026). SCORE likewise uses a saliency-selected parameter subset, not a mask-union construction (Fu et al., 15 Sep 2025).

This comparison suggests a useful taxonomy. MUSE is specifically a coverage-preserving temporal mask downsampling rule for video object removal. It belongs to a wider family of selective-erasure techniques, but its defining contribution is unusually concrete: preserve every target location observed within each temporal compression window so that temporal downsampling does not erase the erase mask itself.