We introduce a selector-extractor framework that extracts high-res features without ever seeing full high-res images to save compute.

We introduce Early Vision Networks (EVNets), hybrid CNNs combining biologically inspired subcortical and cortical front-end blocks, which improve V1 alignment and significantly boost robustness to different image perturbations.

We propose OCTDiff, a bridged diffusion model with Adaptive Noise Aggregation and Multi-Scale Cross-Attention, trained with quality-guided loss. It achieves SOTA enhancement on portable OCT images and shows potential for broader clinical use.

Feed-forward 3D reconstruction with flexibility to rig constraints

We introduce a novel, scalable framework to evaluate compositional generalization, leverage it to evaluate more than 5k models, and propose a family of neural models pushing the Pareto frontier on this task.

Enhancing long video understanding via extreme compression by reducing each selected frame to a single token.

Interpretability methods based on linear, orthogonal features fall short for modern neural representations, which are often hierarchical and nonlinear. Better results come from aligning methods with the true structure of these representations.

We propose a general approach for modifying image-based models to improve robustness and temporal consistency on video.

A self-supervised world model that jointly learns two architecturally distinct representations: one equivariant to specified transformations and another invariant to them

A hybrid quantum-classical Transformer model with quantum-induced doubly-stochastic attention that stabilizes and improves small-scale vision transformers

We present a dataset of floor plan-photo pairs from the Internet with pixel correspondences and camera poses, adapt DUSt3R to improve correspondence prediction, and identify systematic errors for future work.

We propose FACE, a concept-based explanation framework that ensures faithful alignment between model predictions and concept-based representations by combining Non-negative Matrix Factorization with KL divergence regularization.

We enhance diffusion models to be able to recall content from long back in a sequence in order to produce consistent content

WildCAT3D learns multi-view diffusion from images in the wild by explicitly modeling global appearance variations.

InFlux is the first real-world benchmark that provides per-frame ground truth camera intrinsics for videos with dynamic intrinsics, and current baselines struggle to predict accurate intrinsics on our benchmark.

A surprisingly simple plug-and-play method to strengthen adversarial image protection against diverse purification techniques.

To resolve long video understanding task, we propose Deep Video Discovery agent to iterative reasoning and gather information from video content via an agentic search and tool use strategy..

We introduce a directional consistency-driven adaptive density control for 3D Gaussian Splatting to guide split decisions and placements based on structural complexity, improving reconstruction fidelity with fewer primitives.

We introduce Cycle-Sync, a robust and global framework for estimating camera poses (rotations and locations).

We propose a pixel-level black-box attack method to fool STR models to predict more incorrect characters, using a novel multi-population coevolution search algorithm.

Make existing 3D generative models orientation aligned for orientation estimation and efficient insertion.

We propose GLVD, a fast and accurate hybrid method for 3D face reconstruction that combines per-vertex neural fields with dynamically predicted 3D landmarks to achieve state-of-the-art quality from just one or a few images.

Use ground-based cameras to retrieve physical properties of clouds at a high spatial and temporal resolution.

We propose CDC-GS, a 3DGS densification strategy based on visual prior which is perpendicular to existing gradient-based ones. Our method improves the novel view synthesis quality by a large margin without increasing the number of Gaussians.

We implement a framework for learning constraints in vision models and algorithms to support this. We showcase it in Autonomous Driving.

Joint alignment of multiple images without a reference in 50 seconds or less

This paper argues that privacy must be measured and integrated as a core dimension in the evaluation of human pose estimation systems.