Slot Attention, an approach that binds different objects in a scene to a set of "slots", has become a leading method in unsupervised object-centric learning. Most methods assume a fixed slot count K, and to better accommodate the dynamic nature of object cardinality, a few works have explored K-adaptive variants. However, existing K-adaptive methods still suffer from two limitations. First, they do not explicitly constrain slot-binding quality, so low-quality slots lead to ambiguous feature attribution. Second, adding a slot-count penalty to the reconstruction objective creates conflicting optimization goals between reducing the number of active slots and maintaining reconstruction fidelity. As a result, they still lag significantly behind strong K-fixed baselines. To address these challenges, we propose Quality-Guided K-Adaptive Slot Attention (QASA). First, we decouple slot selection from reconstruction, eliminating the mutual constraints between the two objectives. Then, we propose an unsupervised Slot-Quality metric to assess per-slot quality, providing a principled signal for fine-grained slot--object binding. Based on this metric, we design a Quality-Guided Slot Selection scheme that dynamically selects a subset of high-quality slots and feeds them into our newly designed gated decoder for reconstruction during training. At inference, token-wise competition on slot attention yields a K-adaptive outcome. Experiments show that QASA substantially outperforms existing K-adaptive methods on both real and synthetic datasets. Moreover, on real-world datasets QASA surpasses K-fixed methods.
Figures & tables
Figure 1: mBOi vs. slot count K on COCO [ 25 ] . K -fixed methods, such as SPOT [ 20 ] and DINOSAUR [ 33 ] , show large performance fluctuations as K varies. By contrast, K -adaptive methods do not rely on a carefully tuned K . Our method substantially outperforms existing K -adaptive baselines, including MetaSlot [ 26 ] and AdaSlot [ 16 ] , and even surpasses strong K -fixed baselines.
Figure 2: Framework of Quality-Guided K -Adaptive Slot Attention (QASA). The overall training pipeline follows the canonical Slot Attention architecture: an image is encoded by the encoder, fed into the Slot Attention module, and then decoded; the reconstruction loss is computed against embeddings produced by a frozen encoder. The key difference is that we compute a slot-quality metric from the Slot Attention maps Aslot and use it to select a subset of slots, allowing only the selected slots to participate in decoding. At inference, no slot selection is applied. All slots compete via Aslot for each token, yielding a K -adaptive result.
Figure 3: Visualizations on COCO. The MetaSlot examples of AdaSlot [ 16 ] and MetaSlot [ 26 ] are generated using the released checkpoint. ”SPOT-7” and ”SPOT-20” denote SPOT [ 20 ] with K=7 and K=20 , respectively. The SPOT-7 examples are generated using the released checkpoint. The SPOT-20 examples are generated by training a model under the same settings as in the original paper.
COCO
VOC
MOVi-C
MOVi-E
Method
Dec.
K -type
mBOi
mIoU
mBOi
mIoU
mBOi
mBOi
AdaSlot [ 16 ]
MLP
adaptive
27.4
–
–
–
35.6
29.8
MetaSlot [ 26 ]
MLP
adaptive
29.5
27.9
43.9
42.1
35.0
–
QASA
Trans.
adaptive
36.7
35.0
49.7
47.9
46.9
39.1
QASA MLP
MLP
adaptive
33.6
32.5
47.3
45.5
42.2
34.6
Table 1: Comparison of K -adaptive methods. Results for AdaSlot and MetaSlot are taken from their original papers [ 16 , 26 ] . Dashes indicate results not reported.
COCO
VOC
MOVi-C
MOVi-E
Method
Dec.
K -type
mBOc
mBOi
mIoU
mBOc
mBOi
mIoU
mBOi
mIoU
mBOi
mIoU
SLATE [ 34 ]
–
–
33.6
29.1
–
41.5
35.9
–
39.4
37.8
30.2
28.6
DINOSAUR [ 33 ]
Trans.
fixed
39.7
31.6
–
51.2
44.0
–
42.4
–
–
–
DINOSAUR MLP [ 33 ]
MLP
fixed
30.9
27.7
–
40.9
39.5
–
39.1
–
35.5
–
SPOT [ 20 ]
Trans.
fixed
44.7
35.0
33.0
55.6
48.3
–
47.3
46.7
40.1
39.3
SPOT+FS+RC [ 36 ]
Trans.
fixed
45.3
35.7
–
56.5
49.3
–
49.0
47.8
–
–
Table 2: Comparison with K -fixed methods on COCO [ 25 ] , PASCAL VOC [ 15 ] , and MOVi-C/E [ 17 ] . Results for SLATE, DINOSAUR, DINOSAUR MLP , and SPOT are taken from SPOT [ 20 ] . Results for the remaining methods are taken from their original papers [ 36 ] . Dashes indicate results not reported.
Figure 4: Visualizations on MOVi-C/E. The AdaSlot examples are taken from its original paper [ 16 ] . the SPOT examples are generated by training a model under the same settings as in the original paper.
MLP predictor
Linear predictor
Method
Dec.
K -type
class-top1
class-top5
R2 -X
R2 -Y
class-top1
class-top5
R2 -X
R2 -Y
DINOSAUR [ 33 ]
Trans.
fixed
47.1
75.5
55.4
59.1
39.7
68.0
48.8
52.0
DINOSAUR MLP [ 33 ]
MLP
fixed
46.9
75.4
62.9
63.3
39.1
67.5
59.2
59.9
SPOT [ 20 ]
Trans.
fixed
45.9
72.7
51.9
56.4
45.2
72.1
49.2
53.4
AdaSlot [ 16 ]
MLP
adaptive
49.3
79.0
72.7
69.8
33.9
66.3
58.6
53.3
MetaSlot [ 26 ]
MLP
adaptive
40.9
70.0
55.0
55.3
37.9
66.7
52.0
53.2
Table 3: Object property prediction results. We report classification accuracy and R2 scores using an MLP predictor and a linear predictor. DINOSAUR [ 33 ] and MetaSlot [ 26 ] are trained with their official code. SPOT [ 20 ] and AdaSlot [ 16 ] are evaluated using slots extracted from their official checkpoints. All baselines use the same backbone and evaluation setup as our model.
Table 8
Figure 5: Influence of ρ , τ and μ . All results are evaluated on MOVi-C.
MOVi-C
Kmax
15
18
20
22
24
32
mBOi
44.4
46.5
46.9
46.4
46.3
45.4
mIoU
43.4
45.7
46.1
45.6
45.7
45.0
MOVi-E
Kmax
24
28
30
32
34
64
mBOi
38.3
39.1
39.1
39.3
39.5
38.7
mIoU
36.8
37.8
37.8
38.0
38.3
38.3
Table 6: Effect of Kmax on MOVi-C and MOVi-E [ 17 ] .
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Object number distribution on COCO and MOVi-C/E.
Figure 7: Active slot count distribution.
Figure 8: Slot-Quality distribution on COCO.
Method
Train time
Train space
Eval time
SPOT
5m 27s
15.8 GiB
30 s
Ours
5m 47s
17.2 GiB
32 s
Appendix
Table 7: Cost comparison. We report time and per-GPU memory usage for 1-epoch COCO training on 4x RTX 4090s.
Figure 9: Additional examples of QASA on COCO.
Figure 10: Additional examples of QASA on PASCAL VOC.
Object-Centric Representation Learning studies visual models that organize scenes into distinct objects and their properties, positions, and relations.
Slot-based object-centric learning represents an image as a set of latent slots with a decoder that combines them into an image or features. The decoder specifies how slots are combined into an output, but the slot set is typically fixed: the number of slots is chosen upfront and slots are only refined. This can lead to multiple slots competing for overlapping regions of the same entity rather than focusing on distinct regions. We introduce slot merging: a drop-in, lightweight operation on the slot set that merges overlapping slots during training. We quantify overlap with a Soft-IoU score between slot-attention maps and combine selected pairs via a barycentric update that preserves gradient flow. Merging follows a fixed policy, with the decision threshold inferred from overlap statistics, requiring no additional learnable modules. Integrated into the established feature-reconstruction pipeline of DINOSAUR, the proposed method improves object factorization and mask quality, surpassing other adaptive methods in object discovery and segmentation benchmarks.
Christos Chatzisavvas, Panagiotis Rigas, George Ioannakis +2
Department of Electrical and Computer Engineering, Democritus University of Thrace, Greece · Institute for Language and Speech Processing, Athena Research Center, Greece · Department of Informatics & Telecommunications, National and Kapodistrian University of Athens, Greece +1
Studying attention through the lens of interacting particle dynamics has shown how token clustering can emerge from the underlying dynamics. We extend this perspective to slot attention, a method for object-centric image segmentation and representation learning in which learned components obscure how much of the clustering behaviour is intrinsic to the attention dynamics. We therefore introduce simplified slot attention (SSA), a parameter-free variant whose dynamics are connected to soft k-means clustering and which provides a straightforward mechanistic explanation for the emergence of object-centric representations. On the Pascal VOC dataset, SSA achieves performance comparable to that of slot attention, demonstrating that competitive object-centric segmentation can be achieved without learned neural-network components.
Unsupervised video object-centric learning aims to decompose dynamic scenes into persistent, object-level representations without supervision. However, existing slot-based methods struggle to maintain stable object identity in challenging settings such as rapid motion and partial occlusion. First, they typically encode both the per-frame appearance of an object and its identity across frames in a single slot vector, creating an objective conflict that leads to slot swapping: reconstruction requires sensitivity to transient visual changes, whereas temporal consistency requires invariance to them. Second, the token renormalization used in Slot Attention can amplify weakly attending slots, allowing them to absorb tokens from other objects and destabilize slot-to-object correspondence. We propose Dual-State Slot Attention (DSSA), a fully self-supervised framework that addresses these limitations by separating appearance from identity and by reducing spurious updates from weakly matching slots. DSSA decomposes each slot into a local state for per-frame appearance and an identity state for temporally stable object information, thereby aligning reconstruction and temporal consistency with separate representations. The identity state is updated through a learned recurrent transition that acts as a temporal filter on the local state, while competition-modulated aggregation (CMA) down-weights updates from weakly matching slots and prevents them from absorbing tokens from other objects. Experiments on MOVi-C, MOVi-D, and YouTube-VIS demonstrate that DSSA consistently improves segmentation quality and temporal consistency over prior methods, while also yielding stronger downstream object recognition and video dynamics prediction. Code is available at uark-aicv.github.io/dssa.