Computer Vision

Streamlined Inference, a novel training-free framework, dramatically reduces the computation and memory costs of video diffusion models without sacrificing quality, enabling high-resolution video gene…

FashionR2R leverages diffusion models to realistically translate rendered fashion images into photorealistic counterparts, enhancing realism and preserving fine-grained clothing textures.

FairQueue improves fair text-to-image generation by addressing prompt learning’s quality issues through prompt queuing and attention amplification.

FactorizePhys leverages Non-negative Matrix Factorization for a novel multidimensional attention mechanism (FSAM) to improve remote PPG signal extraction from videos.

This paper presents a novel neural network architecture that simultaneously learns to generate and segment images in an unsupervised manner, achieving accurate results across multiple datasets without…

Face2QR: A unified framework generates aesthetically pleasing, scannable QR codes that faithfully preserve facial features, solving the conflict between aesthetics, identity, and scannability.

F-OAL: Forward-only Online Analytic Learning achieves high accuracy and low memory usage in online class incremental learning by using a frozen encoder and recursive least squares to update a linear …

Long-video masked autoencoders (LVMAE) achieve state-of-the-art performance by using an adaptive masking strategy that prioritizes important video tokens, enabling efficient training on 128 frames.

EVA: a novel method generates expressive 3D Gaussian human avatars from monocular RGB videos, excelling in detailed hand and facial expressions via context-aware density control and improved SMPL-X al…

This paper introduces a novel token pruning method for vision state space models, achieving significant computational reduction with minimal performance impact, addressing the limitations of directly …

DUSA:Unlocking Diffusion Models’ Discriminative Power for Efficient Test-Time Adaptation

LOCO Edit achieves precise, localized image editing in diffusion models via a single-step, training-free method leveraging low-dimensional semantic subspaces.

This paper theoretically proves the existence and uniqueness of fixed points in DDIM inversion, optimizing the loss function for improved image editing and extending this approach to unsupervised imag…

FlowDCN: A purely convolutional generative model achieves state-of-the-art image generation speed and quality at arbitrary resolutions, surpassing transformer-based models.

SNELL: Sparse tuning with kerNElized LoRA achieves state-of-the-art parameter-efficient fine-tuning performance with drastically reduced memory usage.

Exo2Ego-V generates realistic egocentric videos from sparse exocentric views, significantly outperforming state-of-the-art methods on a challenging benchmark.

Event-3DGS: First event-based 3D reconstruction using 3D Gaussian splatting, enabling high-quality, efficient, and robust 3D scene reconstruction in challenging real-world conditions.

DSPoser: A novel two-stage approach accurately estimates full-body pose from doubly sparse egocentric video data using masked autoencoders for temporal completion and conditional diffusion models for …

This research introduces adversarial concept preservation, a novel method for safely erasing undesirable concepts from diffusion models, outperforming existing techniques by preserving related sensiti…

Era3D: High-resolution multiview diffusion using efficient row-wise attention, generates high-quality multiview images from single views, overcoming prior limitations.