Geometry-Grounded Diffusion Model Improves Driving Scene View Synthesis
A new arXiv paper proposes GeoV2V, a geometry-grounded video diffusion model aimed at generating consistent driving videos from unseen viewpoints such as lane shifts.
Research Digest
Preprints and papers, rewritten in plain English.
Stable Signer V3: High-Resolution Sign Language Avatars via 4D Splatting
A new arXiv paper proposes a 4D deformable primitive splatting approach for interactive, high-resolution sign language avatars.
Land-Cover-Guided Network Reconstructs MODIS Time Series
A new two-stage network uses land-cover guidance to fill gaps in annual MODIS surface reflectance data, supporting long-term environmental studies.
Visual Scaffolding: Extending Chain-of-Thought to Multimodal Reasoning
A new arXiv paper argues that chain-of-thought reasoning succeeds through textual scaffolding and proposes a visual analogue for multimodal tasks.
Video Reasoning Segmentation: Grounded Reasoning Across Space and Time
A new arXiv paper frames video segmentation as a grounded reasoning problem that requires resolving spatial relations and temporal events.
Image Dehazing Recast as In-Context Tabular Inference
A new arXiv paper proposes reframing image dehazing as a tabular inference problem, potentially avoiding task-specific architectures and extensive training.
HullProbe exposes ship classifiers' reliance on generic hulls
A new probing method measures whether maritime vessel classifiers depend on generic hull features rather than distinctive superstructures.
Sufficient Frame Subsets Aim to Pinpoint Critical Video Moments
A new arXiv paper proposes a way to identify the minimal set of frames that drive a video model's decisions.
ATLAS adapts LVLM attention for reasoning video segmentation
A new arXiv preprint proposes ATLAS, a method that adapts large vision-language model attention for reasoning video object segmentation.
IntriAdapt: using camera intrinsics to supervise test-time adaptation
The method turns camera intrinsics into a supervisory signal for test-time adaptation in metric geometry estimation.
Agentic AI Must Track Design Briefs Over Long Horizons
New research frames the challenge of keeping multimodal AI agents aligned with design intent as a long-horizon problem rooted in design briefs.
MoCKEFA: Model-conditioned keyframes preserve video model beliefs
A new method selects a few keyframes that keep a video model's prediction intact, enabling efficient explanation.
IECLA: A New Approach to Affective Image Editing with Interpretable Emotion Control
A new method, IECLA, aims to make affective image editing more interpretable by learning how different visual elements contribute to a target emotion.
Meta Inverse Prompting Proposed for Video Generative Models
A new arXiv paper introduces meta inverse prompting to address inherent limitations in conditional video generation and editing workflows.
Gaussian Prototypes Boost MRI Super-Resolution at Arbitrary Scales
A new method reconstructs high-resolution 3D MRI volumes from low-resolution anisotropic scans, potentially enabling faster acquisitions without sacrificing image quality.
PRET: A Few-Shot AI System for Pan-Cancer Recognition
A new arXiv preprint describes PRET, a few-shot system that aims to recognize multiple cancer types without requiring example training, potentially easing the global pathologist shortage.
Predicting Lipreading Skill from Eye Movements and Pupils
A new preprint asks how much lipreading ability can be inferred from where and how people look during silent visual speech.
Visual Sufficiency Landscapes Map Which Image Tokens Can Be Pruned
A new arXiv study asks how many subsets of visual tokens can be pruned without changing model behavior, framing reliability as query-dependent.