Semantic Consistency-Guided Masking
- Semantic Consistency-Guided Masking is a strategy that aligns masking processes with semantic structures and invariance constraints instead of using random removals.
- It employs techniques such as connected component extraction, learned semantic parts, and saliency-based selection to enforce coherent, context-preserving predictions.
- This approach is applied across diverse domains including face parsing, masked autoencoding, 3D scene understanding, diffusion-based image generation, and neural machine translation.
Semantic consistency-guided masking denotes a class of training and inference strategies in which masking is coupled to semantic structure, object-level organization, or invariance constraints rather than treated as a purely random removal process. In the surveyed literature, the masked unit may be a connected component in face parsing, a semantic part in masked autoencoding, informative points in large-scale 3D scenes, image patches in semi-supervised segmentation, token-to-region attention in diffusion, words in a bilingual sentence pair, or joints in a skeleton sequence. Across these settings, the central objective is to reduce ambiguity, preserve meaningful context, and enforce coherent predictions or representations under masking (Masi et al., 2019, Li et al., 2022, Xu et al., 2022, Zhou et al., 2023, Cheng et al., 2022, Pan et al., 2023, Wei et al., 18 Aug 2025, Yuan et al., 12 Sep 2025).
1. Conceptual formulation
A recurring premise in this line of work is that conventional random or independent masking is often mismatched to structured data. In face segmentation, pixel-wise softmax plus cross-entropy is criticized for implicitly assuming pixel predictions are i.i.d. and independent, even though faces exhibit strong spatial regularity and occlusions usually form continuous blobs rather than isolated sparse pixels (Masi et al., 2019). In masked autoencoding, SemMAE argues that images lack the semantic decomposition that words provide in NLP, and therefore introduces semantic parts as a visual analogue of words (Li et al., 2022). In large-scale 3D scene understanding, MM-3DScene states that conventional random masking can remove entire objects or distinctive structures, making masked reconstruction highly ambiguous (Xu et al., 2022). In text-to-image diffusion, MaskDiffusion attributes prompt-image mismatch to inadequate cross-modality relation learning and treats unadjusted cross-attention as insufficient for reliable token-to-region alignment (Zhou et al., 2023).
The resulting methods therefore modify one or more of three elements: the unit being masked, the signal used to choose or perturb the mask, and the consistency relation imposed across masked and unmasked views. The mask may be derived from connected components, semantic parts, local statistics, Grad-CAM saliency, weak-to-strong pseudo-labels, or bilingual context. The consistency relation may operate at the level of blob averages, teacher-student pseudo-label agreement, spatial correspondence across progressive masking ratios, diffusion denoising predictions under mask perturbations, or cross-lingually conditioned lexical substitution.
| Domain | Masking unit | Consistency signal |
|---|---|---|
| Face parsing | Blobs / connected components | Blob-average label matching and intra-blob KL agreement |
| MAE / SSL | Semantic parts, informative points, salient joints | Reconstruction under semantic masking |
| Semi-supervised segmentation | Random image patches | Pseudo-label agreement from full-image teacher |
| Diffusion generation | Token-region attention mask or perturbed inpainting mask | Cross-attention alignment or denoising consistency |
| NMT augmentation | Masked words on one side of a parallel pair | Compatibility with local context and paired translation |
This suggests that semantic consistency-guided masking is less a single algorithm than a design principle: masking becomes a means of exposing semantic invariances, suppressing shortcut cues, and reshaping the learning signal around structured units rather than isolated tokens or pixels.
2. Region-consensus supervision in dense prediction
A particularly explicit formulation appears in "Towards Learning Structure via Consensus for Face Segmentation and Parsing" (Masi et al., 2019). That work argues that face segmentation should not be treated as an independent pixel prediction problem, because pixels within the same object should have highly correlated outputs. Its mechanism is blob-level consensus learning. Pseudo-labels are constructed by estimating a full-face mask using 3D face projection, subtracting an existing face segmentation mask to obtain occlusion residuals,
$\boldsymbol{\rho} = \hinge{\bff - \mbf{f}^{occ}, \quad \hinge{\cdot} = \max(0,\cdot),$
refining the residual using morphology, extracting connected components, and merging these components with face/background labels. The resulting semantic classes are defined by
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$
For a blob , the expected probability is
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$
and the loss combines a label-matching term for the blob average with a consistency term that penalizes deviations of individual pixels from that average: $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$ The reported trade-off is . The first term acts as region-level classification; the second enforces consensus within the blob and suppresses noisy, scattered, or isolated pixel predictions. The paper emphasizes fewer artifacts, less sparse masks, and spatially coherent outputs, and evaluates sparsity with
The same work also makes explicit that the approach is not masking in the transformer-style sense. The masking or factorization strategy is instead semantic and morphological: a strong prior full-face contour is combined with a pretrained face mask, occlusion blobs are isolated, and connected components become the units of supervision. In that sense, semantic consistency is guided by blob masks rather than by per-pixel independent labels.
3. Semantic-guided masking in masked representation learning
In self-supervised representation learning, semantic consistency-guided masking commonly appears as a replacement for uniform random masking. "SemMAE: Semantic-Guided Masking for Learning Masked Autoencoders" (Li et al., 2022) first learns semantic parts from a ViT encoder pretrained by iBOT and then uses those parts to drive the masking schedule during MAE pretraining. For an input image , the encoder produces a class token and patch tokens . Part tokens are derived from channel-reweighted class-token features,
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$0
and correlated with patch tokens to produce attention maps
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$1
After reconstruction-based refinement with blur, convolution, and AdaIN, hard part segmentations are obtained by argmax over the attention maps. The masking schedule interpolates between masking a portion of patches within each part and masking whole parts, with
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$2
The paper describes this as an easy-to-hard reconstruction task: first mask patches within each part, then mask all patches of some parts plus some patches of the remaining parts, and finally mask whole parts and predict the rest. Reported results include 84.5% fine-tuning accuracy on ImageNet-1k, 1.4% above vanilla MAE, and a linear-probing result of 68.7 with adaptive masking and $\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$3, compared with 66.8 for random masking and 52.9 for masking 75% parts.
A related but domain-specific design appears in "MaskSem: Semantic-Guided Masking for Learning 3D Hybrid High-Order Motion Representation" (Wei et al., 18 Aug 2025). There the semantic cue is Grad-CAM saliency on relative motion against an average-motion reference,
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$4
with cosine similarity
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$5
Gradients of negative similarity define channel weights and a Grad-CAM score,
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$6
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$7
which is converted to masking probabilities
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$8
Masked indices are then sampled by Gumbel-max. The method deliberately masks the most activated joints, such as hands, arms, head, and other moving parts, in order to prevent shortcut learning. Reconstruction targets are hybrid high-order motion features,
$\forall s,\quad \by(s) = \begin{cases} \text{background}, & \text{if } \bc(s)=0, \ \text{face}, & \text{if } \bc(s)=1, \ \text{occlusion}, & \text{if } \bc(s)\ge 2. \end{cases}$9
with total loss
0
The ablation reports 90.3 for semantic-guided masking versus 89.1 for motion-aware masking on NTU-60 X-view, 90.8 for the hybrid velocity-plus-acceleration target, and best performance at a 90% masking ratio with 1.
These two methods differ in modality and target, but both replace indiscriminate masking with masks aligned to semantically meaningful units. One uses learned semantic parts in images; the other uses saliency over joints and temporal regions in skeleton sequences. In both cases, pure part-level masking from the outset is reported as too hard, while semantic structure combined with a curriculum or weighted target yields better learning.
4. Consistency under masked and progressively masked views
A second major branch uses masking not only to hide information but also to define paired views whose representations or predictions should remain consistent. "MM-3DScene: 3D Scene Understanding by Customizing Masked Modeling with Informative-Preserved Reconstruction and Self-Distilled Consistency" (Xu et al., 2022) argues that applying masked modeling to large-scale 3D scenes is difficult because random masking can make recovery underdetermined. Its response is informative-preserved reconstruction. For each point, local statistics are computed from differences to its neighbors in coordinates and color,
2
Points with high 3 are designated Informative Points and are preserved rather than masked. Masking is progressive: points are sorted by descending local statistics, an incremental masking ratio sequence 4 defines increasingly masked scenes 5, and training reconstructs a more complete scene 6 from a more masked scene 7. The appendix reports that a reconstruction gap of 8 works best. Reconstruction uses a multi-scale symmetric Chamfer distance, while consistency is imposed by an EMA target encoder,
9
and a correspondence-based distillation loss
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$0
The full objective is
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$1
with $\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$2 reported as best. On S3DIS semantic segmentation, informative-preserved masked modeling gives 71.1% mIoU versus 70.4% from scratch, consistency only gives 70.9%, and the combined method reaches 71.9% mIoU and 78.0% mAcc.
"Semi-supervised Semantic Segmentation Meets Masked Modeling: Fine-grained Locality Learning Matters in Consistency Regularization" (Pan et al., 2023) uses a different consistency relation. MaskMatch starts from the weak-to-strong teacher-student framework with EMA update
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$3
and adds a masked modeling proxy task called local consistency regularization. A mask
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$4
is generated by splitting the image into non-overlapping $\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$5 patches and masking each patch independently with probability $\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$6, and the masked input is
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$7
The student predicts segmentation labels on the masked image, but the target remains the teacher’s pseudo-label from the complete image: $\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$8 The paper explicitly notes that even with 30% only visible patches, the student is required to infer full segmentation. A multi-scale ensembling strategy averages teacher predictions over three scales to obtain more reliable pseudo-labels. The combined objective is
$\widehat{\bp}_{\mathbf c} = \frac{1}{|\mathbf c|}\sum_{s\in \mathbf c}\bp(s),$9
In both methods, semantic consistency is enforced across masked views rather than solely within a single masked input. MM-3DScene aligns representations of corresponding visible structures across progressive masking ratios; MaskMatch aligns dense semantic predictions between a teacher that sees the full image and a student that sees only partial evidence. The former focuses on masking-invariant scene representation; the latter on fine-grained locality learning for dense prediction.
5. Conditional masking in diffusion generation and object removal
In generative modeling, semantic consistency-guided masking can operate directly on cross-attention or on the conditioning mask itself. "MaskDiffusion: Boosting Text-to-Image Consistency with Conditional Mask" (Zhou et al., 2023) argues that prompt-image mismatch arises from overlapping attentive regions, preempted attentive regions, and wrong attentive regions. It therefore augments cross-attention with an adaptive conditional mask. Standard attention is written as
$\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$0
with attention map $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$1, and the masked version becomes
$\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$2
The mask is conditioned on the current cross-attention maps and prompt embeddings. Nouns and adjectives are extracted using SpaCy, regions are selected by maximizing total attention while penalizing overlap, and a practical approximation uses top-$\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$3 attention values together with a Gaussian filter and a threshold of 0.5 times the maximum attention value. The algorithm initializes a zeroed mask $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$4 with $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$5 and adds this constant weight to selected token-specific pixels. Temporal momentum,
$\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$6
with $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$7 and $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$8, stabilizes regions across denoising steps. The method is training-free and hot-pluggable, is built on Stable Diffusion v1.4, modifies only cross-attention blocks during inference, and applies masks only at $\frac{1}{|N_C|}\sum_{\mathbf c\in \bc} \Big[ \alpha\,D_{KL}\big(\mathbf y \,\|\, \widehat{\bp}_{\mathbf c}\big) + \frac{\beta}{|\mathbf c|}\sum_{s\in \mathbf c} D_{KL}\big(\widehat{\bp}_{\mathbf c}\,\|\, \bp(s)\big) \Big].$9 feature maps. Reported results are 81s inference time, 83.16% simple prompt support, 74.77% complex prompt support, and CLIP 26.82, compared with 81s, 3.10%, 9.16%, and 26.16 for Stable Diffusion.
"Mask Consistency Regularization in Object Removal" (Yuan et al., 12 Sep 2025) addresses a different generative failure mode: mask hallucination and mask-shape bias in diffusion-based inpainting. Using an SDXL backbone with a plug-and-play ControlNet conditioning module, frozen SDXL parameters, and trainable ControlNet, it defines the standard denoising objective
0
The key contribution is to perturb the inpainting mask while requiring prediction consistency. Two perturbations are used. The first is mask dilation,
1
which expands the mask and encourages broader contextual alignment. The second is mask reshape, implemented either as a minimum enclosing rectangle or as a random reshape
2
with rectangular reshape and random reshape used with 50% probability each. Let 3, 4, and 5. The consistency loss is
6
and the total objective is
7
with 8. On BrushBench, the method reports 60.89 FID, 23.54 PSNR, 0.8969 SSIM, 0.1218 LPIPS, and 0.0741 CMMD. On the 300-sample removal test, it reports 30.35 FID, 29.69 PSNR, 0.8666 SSIM, 0.0835 LPIPS, and 0.0428 CMMD. Ablations show that using both dilation and reshape is better than either alone.
Taken together, these two diffusion methods demonstrate two distinct meanings of mask consistency. MaskDiffusion uses masks to disambiguate token-region correspondence during inference; MCR uses perturbations of the conditioning mask during training so that denoising predictions become invariant to semantically reasonable mask changes. In one case the mask sharpens semantic allocation; in the other it regularizes against over-reliance on mask geometry.
6. Cross-lingual conditional masking in neural machine translation
Semantic consistency-guided masking is not restricted to vision. "Semantically Consistent Data Augmentation for Neural Machine Translation via Conditional Masked LLM" (Cheng et al., 2022) defines semantic consistency for augmented parallel data as two simultaneous requirements: both sides must be fluent in their own language, and the target must remain a high-quality translation of the source. Its mechanism is a Conditional Masked LLM built on BERT and fine-tuned to predict masked tokens on one side of a parallel pair while conditioning on the full other side. For 9, the source-side and target-side conditional objectives are
0
The masking procedure is asymmetric by design: only one side is altered at a time. During CMLM training, 15% of the words on one side are randomly masked while the other side remains fully visible. At augmentation time, replacements are not committed as hard tokens in the final training loop; instead, with probability 1, the original embedding is replaced by a soft embedding derived from the CMLM distribution,
2
The paper identifies this as within-language consistency, because the replacement must fit the local sentence context using both left and right context, and cross-language consistency, because the replacement must remain compatible with the paired translation.
The method is evaluated on IWSLT14 German 3 English, Spanish 4 English, Hebrew 5 English, and WMT14 English 6 German. Reported BLEU improvements are baseline 34.37 to 35.93 on IWSLT De-En, 41.67 to 42.92 on IWSLT Es-En, 33.76 to 35.21 on IWSLT He-En, and 28.25 to 30.15 on WMT14 En-De, the latter corresponding to +1.90 BLEU. On semantic-consistency evaluation for masked-word prediction accuracy on IWSLT14 De-En validation, MLM gives 53.5% on the source and 44.0% on the target, XLM gives 74.8% and 70.4%, and CMLM gives 80.1% and 75.5%. The best BLEU is reported at 7.
This formulation broadens the concept of masking beyond spatial occlusion. The mask here marks lexical positions whose replacements must remain consistent with a bilingual semantic anchor. The consistency constraint is therefore conditional adequacy rather than geometric coherence, but the underlying principle is the same: masking becomes a controlled perturbation whose replacement distribution is constrained by semantics.
7. Recurrent design patterns, misconceptions, and scope
The surveyed methods repeatedly reject the view that masking is valuable simply because it removes information. Several papers instead argue that the benefit of masking depends on whether the removed content corresponds to the right structural unit and whether the model is required to preserve the right invariance. SemMAE reports that masking 75% parts yields 52.9, far below adaptive semantic-guided masking at 68.7, indicating that whole-part masking alone can be too hard (Li et al., 2022). MM-3DScene shows that random masking without progressive reconstruction performs no better than scratch in the adapted PointMAE setup, whereas informative-preserved masking and consistency learning improve downstream performance (Xu et al., 2022). MaskMatch reports that label prediction on masked or unmasked regions works better than image reconstruction and that random patch masking works better than block-wise or grid-wise masking (Pan et al., 2023). MCR identifies exact mask geometry itself as a source of bias, since the model may fill the masked area with an object that mimics the mask’s shape rather than surrounding content (Yuan et al., 12 Sep 2025).
A related misconception is that semantic consistency-guided masking must denote a single masking formalism. The literature instead contains at least four distinct mechanisms. First, masking can define semantically grouped supervision units, as in blob-level connected components for face parsing (Masi et al., 2019). Second, masking can be guided by learned semantic structure or saliency, as in semantic parts, informative points, or Grad-CAM-selected joints (Li et al., 2022, Xu et al., 2022, Wei et al., 18 Aug 2025). Third, masking can define paired views for teacher-student or EMA-based invariance, as in MM-3DScene and MaskMatch (Xu et al., 2022, Pan et al., 2023). Fourth, masking can intervene inside a generative attention mechanism or inpainting condition, as in MaskDiffusion and MCR (Zhou et al., 2023, Yuan et al., 12 Sep 2025).
A plausible implication is that the phrase “semantic consistency-guided masking” is most precise when it refers to the joint design of three components: a semantically meaningful mask construction rule, a target that is compatible with the downstream task, and an explicit consistency criterion across pixels, parts, points, tokens, or mask perturbations. The surveyed papers also indicate that the most effective formulations often avoid two extremes: purely random masking, which may ignore structure, and overly aggressive semantic masking from the outset, which may make the pretext task ill-posed. Progressive schedules, multi-scale pseudo-labeling, momentum stabilization, EMA teachers, and mixed objectives are recurrent responses to that tension.
Within those limits, semantic consistency-guided masking has been instantiated across face parsing, masked autoencoding, 3D scene understanding, semi-supervised segmentation, diffusion-based generation, object removal, neural machine translation, and skeleton-based action recognition. The common claim is not that masking alone yields better models, but that masks become substantially more useful when they are aligned with semantic units and coupled to a consistency objective that constrains what should remain invariant under structured information removal.