Google’s newly unveiled “Banana” visual Transformer introduces dynamic sparse attention that cuts inference cost 3‑5×, reduces memory by 70%, and improves ImageNet accuracy, while demonstrating real‑world gains in autonomous driving, medical imaging, and satellite analysis.
MacTok introduces semantic masking and dual‑space alignment to prevent posterior collapse in continuous image tokenizers, enabling high‑quality generation with just 64‑128 tokens and achieving strong gFID scores on ImageNet at 256×256 and 512×512 resolutions.
A Rochester research team automatically relabeled the entire 1.28 M‑image ImageNet training set with multi‑labels using self‑supervised object discovery and a lightweight region classifier, resulting in a pretrained model that raises COCO mAP by 4.2 points and VOC mAP by 2.3 points.
The paper introduces Drifting Models, a new generative paradigm that moves the distribution evolution to the training phase, achieving true one‑step (1‑NFE) generation with state‑of‑the‑art ImageNet FID scores of 1.54 in latent space and 1.61 in pixel space, while eliminating the need for distillation or classifier‑free guidance.
The new Latent Forcing technique from Fei‑Fei Li’s team reorders the diffusion trajectory, first generating a latent structural sketch and then refining pixel details, which restores efficiency of latent‑space models while preserving 100 % pixel fidelity, achieving state‑of‑the‑art FID scores on ImageNet‑256.
The Pixel Mean Flow (pMF) method eliminates multi‑step sampling and latent‑space encoding, generating high‑quality images in a single step and achieving state‑of‑the‑art FID scores on ImageNet while drastically reducing computational cost.
This paper systematically evaluates 23 transferable adversarial attacks and 11 defenses on ImageNet, revealing that early methods like DI outperform many newer attacks when hyper‑parameters are fairly matched, that diffusion‑based defenses give a false sense of security, and that higher transferability often comes at the cost of reduced stealthiness.
This paper systematically evaluates 23 transferable adversarial attacks and 11 defenses on ImageNet, revealing that early DI attacks surpass newer methods when hyper‑parameters are fairly set, diffusion defenses offer false security, and higher transferability often reduces stealthiness, urging fair benchmarking and comprehensive metrics.
The Mamba-Adaptor introduces a dual‑module adapter that overcomes causal computation limits, long‑range memory decay, and spatial structure loss in state‑space models, delivering state‑of‑the‑art results on ImageNet, COCO, and various downstream visual tasks with minimal overhead.
The article presents MeanFlow, a novel one‑step generative modeling framework that replaces instantaneous velocity with an average‑velocity field, achieving a record‑low FID of 3.43 on ImageNet 256×256 with a single function evaluation and outperforming both prior single‑step and multi‑step diffusion models.
Midjourney enhances its AI models by integrating a wide range of data sources—including public datasets like ImageNet and COCO, academic research from top conferences, partner collaborations, and its own proprietary data—while continuously updating and managing these datasets for quality, privacy, and security.
Midjourney leverages a wide range of data sources—including public datasets like ImageNet and COCO, academic research from top conferences and journals, partner collaborations, and its own proprietary data—augmented by real‑time feeds from Bright Data, to continuously improve and expand its AI models.
The Twins series, a collaboration between Meituan and the University of Adelaide, introduces conditional positional encoding and spatially separable self‑attention to improve efficiency and performance of vision transformers, achieving state‑of‑the‑art results on ImageNet, ADE20K, COCO and high‑precision map segmentation.
JD Research Institute and the University of Sydney introduced ViTAEv2, a 600‑million‑parameter deep learning model that achieved a world‑leading 91.2% top‑1 accuracy on ImageNet Real without external data, demonstrating strong few‑shot learning, reducing labeling costs, and promising advances across many computer‑vision tasks.
Tencent engineers broke the world record for training ImageNet with 128 V100 GPUs in just 2 minutes 31 seconds, detailing a suite of optimizations—including a new Light distributed training framework, single‑machine speed boosts, multi‑machine communication enhancements, and advanced batch convergence techniques—that together dramatically cut training time while maintaining high accuracy.
Fei‑Fei Li, the pioneering AI researcher and former Google Cloud AI lead, rose from humble beginnings in China to create the ImageNet dataset, drive breakthroughs in computer vision, and now returns to Stanford, illustrating how curiosity and perseverance can transform both academia and industry.
Tencent’s intelligent machine‑learning platform achieved a world‑record by training AlexNet in 4 minutes and ResNet‑50 in 6.6 minutes on ImageNet, using large batch sizes, mixed‑precision, LARS optimization, hierarchical synchronization, gradient fusion, and pipeline I/O techniques to overcome accuracy and scalability challenges.
