Computer Vision

FlowTurbo: Blazing-fast, high-quality flow-based image generation via a velocity refiner!

Deep neural networks finally match human biological motion perception capabilities by leveraging patch-level optical flows and innovative neuron designs, achieving a 29% accuracy improvement.

ICTM efficiently solves linear inverse problems using flow priors by iteratively optimizing local MAP objectives, outperforming other flow-based methods.

Biologically-inspired DCnet neural network flexibly modulates visual processing based on context, outperforming existing models on visual search and attention tasks.

A-SAM: Turning SAM’s inherent ambiguity into an advantage for controllable, diverse, and convincing ambiguous object segmentation.

Flatten Anything Model (FAM) revolutionizes neural surface parameterization with unsupervised learning, handling complex topologies and unstructured data fully automatically.

OLIVINE uses visual foundation models for fine-grained image-to-LiDAR contrastive distillation, mitigating self-conflict issues and improving 3D representation learning.

NEMO pinpoints & deactivates neurons memorizing training data in diffusion models, boosting privacy & image diversity.

FIFO-Diffusion generates infinitely long, high-quality videos from text prompts using a pretrained model, solving the challenge of long video generation without retraining.

FFAM uses feature factorization and gradient weighting to produce high-quality visual explanations for 3D object detectors, improving model interpretability and trust.

FewViewGS: A novel method for high-quality novel view synthesis from sparse images using a multi-stage training scheme and a new locality-preserving regularization for 3D Gaussians.

DCOD, a novel two-stage framework (Fetch & Forge), efficiently condenses object detection datasets, achieving comparable performance to full datasets at extremely low compression rates, significantly …

BlackFed: Privacy-preserving federated semantic segmentation using zero/first-order optimization, avoiding gradient/weight sharing!

New feature-level adversarial attacks disrupt visible-infrared person re-identification (VIReID) systems by cleverly aligning and manipulating features to cause incorrect ranking results.

FasterDiT accelerates Diffusion Transformers training 7x without architecture modification by analyzing SNR probability density functions and implementing a new supervision method.

This research dramatically accelerates neighborhood attention, a cost-effective self-attention mechanism, through novel GEMM-based and fused kernel implementations, boosting performance by up to 1759%…

Faster Diffusion achieves significant speedups in diffusion model inference by cleverly reusing encoder features and enabling parallel processing, eliminating the need for computationally expensive di…

FastDrag: One-step image manipulation using generative models, drastically improving editing speed without sacrificing quality.

Conditional Conjugate Integrators (CCI) drastically accelerate sampling in iterative refinement models for inverse problems, achieving high-quality results with only a few steps.

TRACKSTO4D: Fast & accurate 3D reconstruction from casual videos using 2D point tracks, drastically reducing runtime by up to 95% while matching state-of-the-art accuracy.