#01Oct 31, 2026
cs.CV
SASFormer: Stage-Aware Selective Attention Transformer for All-in-One Blind Image Restoration
Xiangpeng Tian, Mingyang Li, Fajing Liu and 3 more
All-in-one image restoration aims to recover high-quality images from multiple degradation types using a single unified model. Recent methods improve unified restoration with hybrid attention, prompt learning, or routing mechanisms, but generally focus on enhancing degradation adaptivity while employing similar feature modeling strategies across network stages. Consequently, the distinct representation requirements of encoding and decoding under diverse degradations remain insufficiently explored. In this paper, we propose SASFormer, a Stage-Aware Selective Attention Transformer for all-in-one blind image restoration. SASFormer follows a stage-aware design principle: the encoder uses established hybrid channel-spatial attention to build robust content-preserving and degradation-aware representations, while the decoder employs adaptive selective attention for restoration-oriented reconstruction. To address the stage-specific requirements of decoder reconstruction, we develop Adaptive Selective Channel Self-Attention (ASCSA) and Adaptive Selective Spatial Self-Attention (ASSSA), enabling the decoder to dynamically identify restoration-relevant channel responses and spatial contexts. Specifically, ASCSA performs adaptive query-channel selection and attention masking, while ASSSA adaptively integrates spatial attention across multiple window sizes. Compact implicit prompts further modulate the selection process without requiring explicit degradation labels. Extensive experiments on 3-, 5-, and 11-degradation settings demonstrate that SASFormer achieves strong restoration performance with favorable parameter efficiency. Moreover, stage-wise representation analysis and ablation studies provide empirical support for the proposed stage-aware design and the effectiveness of decoder-side adaptive selection.
#02Oct 30, 2026
cs.CV
PRISM-Net: Dynamic Routing Network with VLM-Guided Disentangled Attribute Distillation for Explainable Fine-Grained Product Retrieval
Chung-Ming Lo, Che-Wei Chang
Fine-grained visual product retrieval in large-scale e-commerce multimedia systems demands discriminative representations capable of simultaneously resolving macro-level category semantics and micro-level visual attributes such as structural contour (line) and chromatic appearance (color). However, conventional backbones including convolutional neural networks (CNNs) and vision Transformers (ViTs) inherently suffer from feature entanglement and shortcut learning. To learn the micro-level visual attributes without costly manual attribute annotations, we propose phase-routed interpretable semantic mixture-of-experts network (PRISM-Net), an explainable multi-expert architecture that disentangles holistic visual features into disentangled line and color expert pathways governed by a conditional meta-router. In phase 1, we introduce an external VLM-Guided sigmoid semantic distillation mechanism with gradient-detached routing gates, utilizing zero-shot text prompts from SigLIP to supervise specialized structural and chromatic experts. In phase 2, the disentangled expert embeddings are adaptively fused with macro base features via a learned meta-router gate to calibrate category boundaries. Extensive 5-fold cross-validation across 275,454 e-commerce images (18 categories) and full-scale retrieval benchmarks (55,091 queries against a 220,363-image gallery) across SwinV2, ConvNeXt-V2, and DINOv3 demonstrate substantial gains. On SwinV2, PRISM-Net boosts full-fold product mAP@10 from 80.98% to 83.90% (+2.92%), line mAP@10 by +5.74%, and color mAP@10 by +9.12%. Furthermore, unsupervised distribution analysis of the meta-router gate reveals that both ViT and CNN backbones emergently allocate up to 6.8x higher structural line weights to rigid appliances and hardware while dedicating >96% chromatic weight to clothing.
#03Oct 31, 2026
cs.CV
Words Before Pixels: Selective Modality Routing for Vision-Language Model Unlearning
Laura Yao, Haochen Zhang, Jinhao Duan and 2 more
Vision-language model (VLM) unlearning is often treated as an objective-design problem: given multimodal forget data, the goal is to optimize a loss that removes undesirable behavior while preserving retained capabilities. We argue that this view overlooks an equally important data-centric question: how should each forget example be represented before optimization? Existing methods typically keep image-conditioned failures in their original image-text form, implicitly assuming that pixels are the right route for unlearning whenever pixels appear in the input. We challenge this assumption by constructing modality-decomposed versions of VLM benchmarks, enabling controlled comparisons among image-only, text-only, and multimodal representations of the same examples. Our modality-mixing experiments show that unsafe instruction-following behavior is more effectively suppressed through textual representations than through images alone, while purely text-only unlearning can still weaken visual grounding when the target behavior depends on visual evidence. Motivated by this trade-off, we introduce ModRoute, a gradient-guided per-sample modality routing strategy that selectively converts high visual-pressure examples to text-only form while keeping the remaining examples multimodal. Notably, on VLGuard with LLaVA-1.5-7B, ModRoute has the strongest composite unlearning-utility score of 0.847 at $γ=0.5$, a 2.5% improvement over the best random-switching score and $\geq$46% better than pure multimodal or text unlearning. Overall, our results show that effective VLM unlearning depends not only on the objective, but also on the composition and per-sample modality routing of the unlearning data.
#04Oct 31, 2026
cs.CV
Selecting or Solving? What Best-of-N Verification Delivers in Video Reasoning
Martin Q. Ma, Yuxiao Qu, Ruslan Salakhutdinov and 3 more
Answering a complex question about a long video is a grounding problem: the answer turns on a few seconds of evidence, and vision-language models remain well below human accuracy at finding them. The standard way to spend extra test-time compute on this kind of video reasoning is best-of-N selection: sample several chain-of-thought answers from a fixed model, then let a second model re-watch the video and pick one. Such pipelines are reported by a single headline number, their margin over a majority vote. That number conflates two effects with opposite deployment consequences: the benefit the second model adds by selecting among candidates, and the benefit it would have supplied by solving the question itself. We separate them, measuring the Selection Increment (SI) -- selection accuracy minus the same verifier's accuracy with the candidates withheld -- alongside the margin over self-consistency, across five long-video benchmarks. The Selection Increment is largest for the weakest verifiers: a verifier that solves poorly on its own leaves the candidates the most room to help. Beating the majority vote follows the opposite profile: where the vote already sits near its ceiling no verifier we test improves on it; below that ceiling only the strongest do. On Video-Holmes, the two come apart on a single pool of candidates -- a weak verifier gains +19.3 points from the candidates yet lands 7.1 points below the vote, while the frontier verifier gains the least and beats the vote. The weakest verifier also shows a failure mode: on EgoSchema it picks by where an option sits in the list rather than by what the video shows. Read as a merit score, SI puts a verifier that fails to beat a majority vote at the top of its ranking; read as an audit, it says which of the two effects a headline number delivers. Code is available at https://github.com/facebookresearch/best-n-selection-video-reasoning.
#05Oct 31, 2026
cs.CV
3D Neural Rendering for Automated Ultrasonic Laser Scans
Thomas Pierson, Matthew Mitchell, Kira Perna and 3 more
Laser doppler vibrometer (LDV) scans can be used to detect defects and cracks in manufactured parts before they are deployed. This work presents a method for reconstructing the steady-state surface wavefield of a three-dimensional part by combining automated robotic manipulation with neural rendering. A six-degree-of-freedom (6-DOF) robotic manipulator (UFactory xArm6) with an embedded piezoelectric transducer presented an acoustically excited test article to a stationary LDV at a sequence of known orientations, yielding a set of complex-valued 2D scans with poses derived analytically from the arm kinematics rather than from image-based pose estimation. These scans were used to train a novel complex-nerfacto model, a modified Neural Radiance Field (NeRF) pipeline that replaces the standard three-channel RGB output head with a two-channel output. Complex-nerfacto represents the real and imaginary velocity components of a wavefield, preserving the complex measurement without quantization. Reconstructions of a welded aluminum cylinder and a welded T-bracket were evaluated against measured scans using Z-normalized cross-correlation (ZNCC). The complex-valued pipeline achieved a mean ZNCC of 0.60 at held-out viewing angles, compared with 0.28 for a baseline RGB nerfacto model trained on rendered images of the same data. The resulting surface wavefield reconstructions supported defect detection and characterization on complex components. We identify geometry reconstruction from sparse LDV scans as the remaining obstacle to practical deployment.