• Video Diffusion Solves "Fast" And "Good" As One Problem. Reward Lightning optimizes preference alignment and distillation acceleration on a shared latent backbone, generating in 1 to 4 steps and lifting the VBench average by 2.1%.

• Concept Unlearning Picks The Worst Prompt First, Then Fine-Tunes. MPSelectTune treats prompt shape itself as a knob for unlearning robustness. Training only on the worst prompt holds up better than mixing all prompt types.

• Industrial Inspection Takes A Verifiable, Training-Free Route. GLLS grounds text criteria to visual evidence with a dual-stream in-context design, dodging the cold-start problem of having no defect samples. Every diagnosis is traceable.

• Sparse-View Reconstruction Fixes Its Geometry Weakness. High-confidence depth plus a normal prior fills the gaps, beating the previous best sparse-view surface reconstruction on DTU and Tanks-and-Temples.

Also Notable

• Amazon Estimates Rainforest Biomass From Three Modalities — Optical plus PolInSAR with a phase-preserving Transformer, dodging SAR signal saturation and cloud cover.
• Video Segmentation Tracks Instances By Geometric Grounding — Dropping the appearance memory bank makes object consistency across views and time steadier.
• Audio-Visual QA Uses Question-Guided Tri-Modal Attention — Avoids the information loss from deep stacks of self and cross attention.
• City-Scale 3DGS Surface Reconstruction — View-orientation partitioning plus scene completion tackles large-scene geometry.
• Order-Equivariant Networks Unify Graph And Sheaf — Extends equivariant deep learning to richer symmetry structures.
• Articulated 3D Reconstruction Bootstrapped From 2D Image Sets — Drops the dependence on hand-made synthetic paired data.
• Conditional Image Generation Wants Both Ends — Task-specific performance without crushing rich conditions into one scalar objective.
• Action And Gaze Modeled Jointly In Sync — Reads human interaction intent instead of computing each separately.

Read the full edition →