All papers

The paper uses EEG data to compare how image-text aligned and vision-only models align with developing human visual representations.

The paper introduces an optical music recognition dataset and pipeline linking medieval chant manuscript images to musical target tokens.

The paper introduces a framework combining targeted patch perturbation with textual explanations to improve face forgery detection generalization.

The paper proposes a training-free guidance method enabling continuous valence-arousal emotional control in foundation image editors.

The paper analyzes camera and LiDAR detection discrepancies across driving datasets to evaluate assumptions in multimodal fusion design.

ConDenseDrive combines knowledge distillation and token pruning to create an efficient visual encoder for end-to-end autonomous driving.

FOCUS is a specialized pathology foundation model pretrained on intraoperative frozen-section slides of central nervous system tissue.

UniMetric provides a training-free framework that recovers metric scale in visual geometry foundation models using object size priors.

DDVR-Net decouples glyph structures from physical damage and uses visibility-guided restoration for historical document OCR.

UniLAP introduces a continuous latent action policy trained via flow matching to enable robot manipulation without relying on future frames.

The paper presents a self-supervised approach using a simple 3D prior to learn semantically consistent visual features for complex objects.

MaskSense addresses decoding exploration biases in masked generative models when aligning them with reinforcement learning.

DIME scales facial representation learning using a differential masked autoencoder that supervises differences between identity-matched image pairs.

PerturbBridge directly supervises perturbation-specific phenotypes during generative training to model cellular responses in images.

The authors present a three-branch deep learning framework for automated aerial wildlife detection, strip delineation, and image matching.

The study compares LLM guidance integration strategies for visual question answering, showing shallow late fusion outperforms deeper attention injection.

The paper introduces a visual-memory framework enabling unified multimodal models to handle long-horizon interleaved text and image generation.

The paper uses contrastive learning with a Siamese network to extract key visual frames and detect video reuse.

EgoProbe benchmarks long-range retention and event-specific recall under interference for streaming egocentric video models.

MARS accelerates world action model inference across transformer layers, diffusion denoising, and policy queries without retraining.

Multics provides a hierarchical communication framework for multi-player interactive video world models to reduce quadratic attention scaling costs.

DexNinja introduces a real-to-sim-to-real framework to train robotic policies for dexterous cutting of deformable objects.

AReA uses an adaptive, coarse-to-fine receptive field attention mechanism to enable efficient high-resolution generation in diffusion transformers.

LIMIT aligns slide-level pathology reports with patch-level vision features by freezing the vision encoder and training only the text encoder.

CameraVLM couples motion states and on-screen flow dynamics to improve camera-motion understanding in vision-language models.

The authors introduce a hybrid architecture combining EfficientNetV2 and Swin Transformer to classify Alzheimer's and Parkinson's disease from brain MRI.

HairFlow generates 3D hair strands using a multimodal diffusion framework conditioned on coarse meshes and multi-view sketches.

WIRE reduces video LLM inference costs by moving video computations from input tokens into temporary low-rank weight updates.

The paper presents a dataset and baseline model to track spatio-temporal human head avatars directly from EEG brain signals.

TableVision recovers verified table structure from historical documents using dynamic programming and column sum checks.

Proposes a post-training decoder alignment method for latent generative models to resolve distribution mismatches between training and generation latents.

Investigates active learning dynamics by regularizing representation collapse and managing sample acquisition across training iterations.

Surveys real-time 3D Gaussian Splatting scene reconstruction methods, systems, and practical trade-offs across five major technical categories.

Proposes a two-phase token pruning strategy for vision-language models and identifies evaluation flaws in common decode-phase benchmarks.

Introduces a cross-species portrait animation framework that decomposes motion into head poses and expression keypoints to prevent driving-species leakage.

Presents a single-pass framework inferring simulation-ready material parameters from multi-view deformation videos using projected physical response fields.

Presents an error-bounded progressive compression codec for 3D Gaussian Splatting with user-declared tolerances for primitive distortion.

Introduces a self-supervised background removal framework for high-dimensional microscopy using local pixel shuffling and noise-reduction techniques.

The paper presents a physically consistent world model incorporating robot-terrain contact dynamics to predict quadruped action outcomes from an RGB image.

Proposes a calibrated disambiguation model to optimize sample weighting when training classifiers on datasets with noisy or excessive labels.

Proposes a lightweight aesthetic cropping model built on a small vision-language model using multi-stage candidate optimization.

Presents an embodied agent framework combining streaming spatial memory with active exploration for visual reasoning in physical environments.

Extends dynamic world generation to produce physics-executable simulations with collision meshes and physical dynamics for embodied agents.

Develops an agentic video reasoning framework that iteratively identifies missing information, gathers visual evidence, and grounds temporal clues.

Introduces an evaluation benchmark to assess user-generated image quality and aesthetics across diverse real-world scenes.

Presents a framework integrating posture neutralization with statistical shape modeling to isolate anatomical variations for ergonomic product design.

Introduces a deficit-driven stratified acquisition algorithm to create a balanced, co-registered satellite imagery corpus for self-supervised Earth observation.

Adapts text-to-video diffusion models for surgical procedure simulation using hierarchical procedural decomposition and progressive curriculum learning.

Introduces a scene-grounded Wasserstein distance objective to optimize graph-level semantic proximity in 3D scene graph prediction tasks.

Proposes a differentiable graph-level loss function to measure semantic and structural discrepancies between predicted 3D scene graphs and ground truth.

Introduces a low-rank adaptation method that splits rank budgets into stable and dynamic branches to capture residual gradients outside the tangent space.

Proposes an unsupervised image retargeting framework combining dense visual attention with mesh warping to preserve content without manual annotations.

Presents a framework that segments long-horizon human manipulation videos into skills and translates them into executable robotic actions.

Introduces an explainability framework and weakly supervised video-text corpus to evaluate causal and temporal facial expression understanding in multimodal models.

This paper introduces a training-free framework combining multi-hypothesis search and local visual verification to ground interface instructions in graphical user interfaces.

This paper proposes a bidirectional polarity-aware cross-modality fusion module to improve multimodal semantic segmentation across diverse complementary sensor streams.

This paper proposes a hierarchical framework for learning whole-arm robot contact manipulation tasks from human demonstrations.

This paper introduces an active learning framework balancing uncertainty with submodular spatial coverage for chest X-ray lesion segmentation.

This paper analyzes how different visual representation types affect the 3D spatial reasoning capabilities of multimodal large language models.

This paper presents a knowledge distillation framework that reconstructs dense teacher visual tokens to train efficient compressed student vision-language models.