Dense2MoE：将扩散Transformer重构为MoE以实现高效文本到图像生成

Residual Mixture of Experts

Mixture of Experts (MoE) is able to scale up vision transformers effectively. However, it requires prohibiting computation resources to train a large MoE transformer. In this paper, we propose Residual Mixture of Experts (RMoE), an efficient training pipeline for MoE vision transformers on downstream tasks, such as segmentation and detection. RMoE achieves comparable results with the upper-bound MoE training, while only introducing minor additional training cost than the lower-bound non-MoE training pipelines. The efficiency is supported by our key observation: the weights of an MoE transforme…

MiM-DiT：基于扩散变换器的MoE在MoE中实现全功能图像修复

MiM-DiT: MoE in MoE with Diffusion Transformers for All-in-One Image Restoration

MiM-DiT通过双层MoE架构与预训练扩散模型，实现对多种图像退化类型的统一修复，提升复杂真实场景下的修复效果。

TAG-MoE：基于任务的门控用于统一的生成混合专家

TAG-MoE: Task-Aware Gating for Unified Generative Mixture-of-Experts

本文提出TAG-MoE框架，通过引入层次化任务语义标注方案和预测对齐正则化，解决密集扩散变换器架构中任务干扰问题，提升生成质量和保真度。

MoECa: 在扩散变换器中对齐特征复用与专家分解

MoECa: Aligning Feature Reuse with Expert Decomposition in Diffusion Transformers

针对DiT-MoE中跨时间步的冗余计算，提出基于专家分支级别的细粒度缓存框架MoECa，实现分支级特征复用，并引入专家感知自适应控制和同步缓存更新，在多个模型上取得高达2.83倍加速且质量损失极小。

Cross-Modal Alignment with Mixture Experts Neural Network for Intral-City Retail Recommendation

In this paper, we introduce Cross-modal Alignment with mixture experts Neural Network (CameNN) recommendation model for intral-city retail industry, which aims to provide fresh foods and groceries retailing within 5 hours delivery service arising for the outbreak of Coronavirus disease (COVID-19) pandemic around the world. We propose CameNN, which is a multi-task model with three tasks including Image to Text Alignment (ITA) task, Text to Image Alignment (TIA) task and CVR prediction task. We use pre-trained BERT to generate the text embedding and pre-trained InceptionV4 to generate image patc…

基于CLIP和MoE变换器的上下文感知注视估计

GazeFormer-MoE: Context-Aware Gaze Estimation via CLIP and MoE Transformer

GazeFormer-MoE通过CLIP和MoE变换器实现上下文感知的注视估计，取得新的最先进角误差，提升64%的相对性能。

MoE中的路由问题：通过显式路由指导扩展扩散Transformer

Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance

ProMoE通过显式路由指导提升视觉MoE性能，结合两步路由策略和原型路由机制，在ImageNet基准上优于现有方法。

EditYourself: 基于音频驱动的谈话头视频生成与操控的扩散变换器

EditYourself: Audio-Driven Generation and Manipulation of Talking Head Videos with Diffusion Transformers

EditYourself通过音频驱动的视频到视频编辑框架，实现基于脚本的谈话头视频修改，包括内容的添加、删除和重新定时，同时保持唇同步和时间一致性。

Mobile V-MoEs：通过稀疏混合专家缩小视觉Transformer

Mobile V-MoEs: Scaling Down Vision Transformers via Sparse Mixture-of-Experts

本文提出一种移动友好的稀疏混合专家设计Mobile V-MoEs，通过整图路由和超类引导的稳定训练，在资源受限场景下使视觉Transformer在性能与效率间取得更优平衡。

Transforming Vision Transformer：迈向高效的多任务异步学习

Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learning

针对视觉Transformer多任务学习中MoE与LoRA结合效果不佳的问题，提出EMTAL方法，通过MoEfied LoRA结构、质量保持优化机制和路由衰减策略，实现高效的多任务异步学习，性能优于现有最优方法。

Patch-MoE Mamba: 一种用于医学图像分割的基于补丁顺序的专家混合状态空间架构

Patch-MoE Mamba: A Patch-Ordered Mixture-of-Experts State Space Architecture for Medical Image Segmentation

本文提出了一种基于补丁顺序的专家混合状态空间架构Patch-MoE Mamba，以解决现有Mamba分割模型在像素级方向扫描破坏局部二维空间结构以及简单求和融合方向无法适应多样物体大小、形状和边界的问题。

YOLO-Master：基于专用变换器的MOE加速以提升实时检测

YOLO-Master: MOE-Accelerated with Specialized Transformers for Enhanced Real-time Detection

YOLO-Master通过引入高效的稀疏专家混合模块，实现了基于场景复杂度的自适应计算，提升了实时目标检测的精度与效率。

MoE-FFD：用于广义且参数高效的人脸伪造检测的专家混合模型

MoE-FFD: Mixture of Experts for Generalized and Parameter-Efficient Face Forgery Detection

针对基于ViT的人脸伪造检测方法存在的全微调资源消耗大、局部线索捕捉弱、泛化性有限等问题，提出MoE-FFD，通过冻结ViT骨干、更新轻量级LoRA和适配器层，结合Transformer表达能力与CNN局部先验，设计MoE模块，实现参数高效且泛化性强的最先进检测性能。

eMoE-Tracker: 基于环境MoE的Transformer用于鲁棒事件引导目标跟踪

eMoE-Tracker: Environmental MoE-based Transformer for Robust Event-guided Object Tracking

针对现有事件-帧多模态跟踪忽略环境属性、模板与搜索特征交互不足的问题，提出含eMoE环境解耦组装模块和CRM对比关系建模模块的eMoE-Tracker框架，在多基准上取得SOTA性能。

GNN-MoE：使用GNN的上下文感知图像块路由实现参数高效的领域泛化

GNN-MoE: Context-Aware Patch Routing using GNNs for Parameter-Efficient Domain Generalization

本文提出GNN-MoE，将图神经网络路由与Kronecker适配器MoE框架结合，利用图像块间关系进行上下文感知路由，以高参数效率实现领域泛化。

CLEAR-MoE: 通过校准驱动的层选择从冻结视觉Transformer中提取共享基专家

CLEAR-MoE: Shared-Basis Expert Extraction from Frozen Vision Transformers via Calibration-Driven Layer Selection

提出CLEAR-MoE四阶段后训练流程，将冻结的ViT转换为稀疏MoE模型，通过共享SVD基保留精度，在多个ViT骨干上精度损失≤0.10%。

专家权重平均：一种面向视觉Transformer的新型通用训练方案

Experts Weights Averaging: A New General Training Scheme for Vision Transformers

该研究提出面向视觉Transformer的专家权重平均通用训练方案，通过训练时引入高效MoE并迭代平均专家权重、推理时还原为原结构，在不增加推理成本的前提下提升性能，且可用于微调及增强朴素MoE效果。

ExFusion：通过多专家融合实现高效的Transformer训练

ExFusion: Efficient Transformer Training via Multi-Experts Fusion

ExFusion通过多专家融合提升Transformer训练效率，在保持性能的同时降低计算成本。

Edge-MoE：基于任务级稀疏性的混合专家内存高效多任务视觉Transformer架构

Edge-MoE: Memory-Efficient Multi-Task Vision Transformer Architecture with Task-level Sparsity via Mixture-of-Experts

Edge-MoE提出首个多任务ViT的FPGA加速器，通过自注意力重排序、单遍softmax和GELU近似、统一计算单元及patch重排序等创新，解决了M$^3$ViT部署挑战，能效较GPU和CPU分别提升2.24倍和4.90倍。

GMGaze：基于MoE的上下文感知 gaze 估计方法结合 CLIP 和多尺度 Transformer

GMGaze: MoE-Based Context-Aware Gaze Estimation with CLIP and Multiscale Transformer

本文提出 GMGaze 方法，通过多尺度 Transformer 和 CLIP 结合，利用语义原型条件和稀疏 MoE 模块提升 gaze 估计精度，实验表明在多个基准上均优于现有方法。

Vision Transformer 可以有多轻量

How Lightweight Can A Vision Transformer Be

本文提出利用 Mixture-of-Experts (MoE) 精简 vision transformers，通过共享参数、深度缩放及分组查询注意力等策略，在仅 0.67M 参数规模下实现了具竞争力的性能。

Scaling Vision with Sparse Mixture of Experts

Sparsely-gated Mixture of Experts networks (MoEs) have demonstrated excellent scalability in Natural Language Processing. In Computer Vision, however, almost all performant networks are "dense", that is, every input is processed by every parameter. We present a Vision MoE (V-MoE), a sparse version of the Vision Transformer, that is scalable and competitive with the largest dense networks. When applied to image recognition, V-MoE matches the performance of state-of-the-art networks, while requiring as little as half of the compute at inference time. Further, we propose an extension to the routi…

3D-MoE：迈向空间智能——基于混合专家模型的3D推理与动作生成

3D-MoE: Towards Spatial Intelligence with Mixture-of-Experts for 3D Reasoning and Action Generation

针对3D VLM计算成本高的问题，提出3D-MoE，采用混合专家与概率路由，集成Pose-DiT动作头，以更少激活参数在3D基准和机器人操作中实现高性能与实时推理。

一种专家混合（MoE）模型，用于在组织病理学图像模糊程度变化的情况下提升基于AI的计算病理学预测性能

A Mixture of Experts (MoE) model to improve AI-based computational pathology prediction performance under variable levels of histopathology image blur

本文提出专家混合（MoE）策略，通过组合不同模糊水平数据训练的专家模型，显著提升组织病理学图像模糊条件下基于AI的预测性能，并在乳腺癌WSI分级和IHC预测任务中验证了其优于基线模型的效果。

Cross-token Modeling with Conditional Computation

Mixture-of-Experts (MoE), a conditional computation architecture, achieved promising performance by scaling local module (i.e. feed-forward network) of transformer. However, scaling the cross-token module (i.e. self-attention) is challenging due to the unstable training. This work proposes Sparse-MLP, an all-MLP model which applies sparsely-activated MLPs to cross-token modeling. Specifically, in each Sparse block of our all-MLP model, we apply two stages of MoE layers: one with MLP experts mixing information within channels along image patch dimension, the other with MLP experts mixing inform…

Patcher：用于精确医学图像分割的专家混合 Patch Transformer

Patcher: Patch Transformers with Mixture of Experts for Precise Medical Image Segmentation

本文提出 Patcher 医学图像分割架构，通过重叠大块内小块 Transformer 编码和专家混合解码器，在卒中病变与息肉分割上显著超越现有方法。

EVE：基于掩码预测与模态感知 MoE 的高效视觉-语言预训练

EVE: Efficient Vision-Language Pre-training with Masked Prediction and Modality-Aware MoE

本文提出统一多模态 Transformer 基础模型 EVE，利用模态感知 MoE 与图像-文本掩码信号建模实现高效预训练，训练提速 3.5 倍并在多项视觉-语言任务上达到 SOTA。

超越自注意力：用于快速图像描述的次二次视觉Transformer

Beyond Self-Attention: Sub-Quadratic Vision Transformers for Fast Image Captioning

提出基于高斯混合模型和EM算法的概率Transformer，将自注意力复杂度从二次降至线性，在Flickr30K上实现高效图像描述。

Life-IQA: 通过GCN增强的层交互和MoE基于的特征解耦提升盲图像质量评估

Life-IQA: Boosting Blind Image Quality Assessment through GCN-enhanced Layer Interaction and MoE-based Feature Decoupling

Life-IQA通过GCN增强的层交互和MoE基于的特征解耦提升盲图像质量评估的准确性和效率。