EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing

• URL: http://arxiv.org/abs/2410.02098v5
• Date: Tue, 04 Mar 2025 17:23:51 GMT
• Title: EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing
• Authors: Haotian Sun, Tao Lei, Bowen Zhang, Yanghao Li, Haoshuo Huang, Ruoming Pang, Bo Dai, Nan Du,
• Abstract summary: We develop a new family of Mixture-of-Experts (MoE) models (EC-DIT) for diffusion transformers with expert-choice routing.<n>EC-DIT learns to adaptively optimize the compute allocated to understand the input texts and generate the respective image patches.<n>In text-to-image alignment evaluation, our largest models achieve a state-of-the-art GenEval score of 71.68% and still maintain competitive inference speed with intuitive interpretability.
• Score: 43.63139112740413
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Diffusion transformers have been widely adopted for text-to-image synthesis. While scaling these models up to billions of parameters shows promise, the effectiveness of scaling beyond current sizes remains underexplored and challenging. By explicitly exploiting the computational heterogeneity of image generations, we develop a new family of Mixture-of-Experts (MoE) models (EC-DIT) for diffusion transformers with expert-choice routing. EC-DIT learns to adaptively optimize the compute allocated to understand the input texts and generate the respective image patches, enabling heterogeneous computation aligned with varying text-image complexities. This heterogeneity provides an efficient way of scaling EC-DIT up to 97 billion parameters and achieving significant improvements in training convergence, text-to-image alignment, and overall generation quality over dense models and conventional MoE models. Through extensive ablations, we show that EC-DIT demonstrates superior scalability and adaptive compute allocation by recognizing varying textual importance through end-to-end training. Notably, in text-to-image alignment evaluation, our largest models achieve a state-of-the-art GenEval score of 71.68% and still maintain competitive inference speed with intuitive interpretability.

Related papers

• Boosting Generative Image Modeling via Joint Image-Feature Synthesis [10.32324138962724]
  We introduce a novel generative image modeling framework that seamlessly bridges the gap by leveraging a diffusion model to jointly model low-level image latents. Our latent-semantic diffusion approach learns to generate coherent image-feature pairs from pure noise. By eliminating the need for complex distillation objectives, our unified design simplifies training and unlocks a powerful new inference strategy: Representation Guidance.
  arXiv  Detail & Related papers  (2025-04-22T17:41:42Z)
• Adversarial Semantic Augmentation for Training Generative Adversarial Networks under Limited Data [27.27230943686822]
  We propose an adversarial semantic augmentation (ASA) technique to enlarge the training data at the semantic level instead of the image level. Our method consistently improve the synthesis quality under various data regimes.
  arXiv  Detail & Related papers  (2025-02-02T13:50:38Z)
• FiT: Flexible Vision Transformer for Diffusion Model [81.85667773832279]
  We present a transformer architecture specifically designed for generating images with unrestricted resolutions and aspect ratios. Unlike traditional methods that perceive images as static-resolution grids, FiT conceptualizes images as sequences of dynamically-sized tokens. Comprehensive experiments demonstrate the exceptional performance of FiT across a broad range of resolutions.
  arXiv  Detail & Related papers  (2024-02-19T18:59:07Z)
• AICT: An Adaptive Image Compression Transformer [18.05997169440533]
  We propose a more straightforward yet effective Tranformer-based channel-wise auto-regressive prior model, resulting in an absolute image compression transformer (ICT) The proposed ICT can capture both global and local contexts from the latent representations. We leverage a learnable scaling module with a sandwich ConvNeXt-based pre/post-processor to accurately extract more compact latent representation.
  arXiv  Detail & Related papers  (2023-07-12T11:32:02Z)
• StraIT: Non-autoregressive Generation with Stratified Image Transformer [63.158996766036736]
  Stratified Image Transformer(StraIT) is a pure non-autoregressive(NAR) generative model. Our experiments demonstrate that StraIT significantly improves NAR generation and out-performs existing DMs and AR methods.
  arXiv  Detail & Related papers  (2023-03-01T18:59:33Z)
• TcGAN: Semantic-Aware and Structure-Preserved GANs with Individual Vision Transformer for Fast Arbitrary One-Shot Image Generation [11.207512995742999]
  One-shot image generation (OSG) with generative adversarial networks that learn from the internal patches of a given image has attracted world wide attention. We propose a novel structure-preserved method TcGAN with individual vision transformer to overcome the shortcomings of the existing one-shot image generation methods.
  arXiv  Detail & Related papers  (2023-02-16T03:05:59Z)
• Factorizer: A Scalable Interpretable Approach to Context Modeling for Medical Image Segmentation [6.030648996110607]
  This work introduces a family of models, dubbed Factorizer, which leverages the power of low-rank matrix factorization for constructing an end-to-end segmentation model. Specifically, we propose a linearly scalable approach to context modeling, formulating Nonnegative Matrix Factorization (NMF) as a differentiable layer integrated into a U-shaped architecture. Factorizers compete favorably with CNNs and Transformers in terms of accuracy, scalability, and interpretability.
  arXiv  Detail & Related papers  (2022-02-24T18:51:19Z)
• ViTAEv2: Vision Transformer Advanced by Exploring Inductive Bias for Image Recognition and Beyond [76.35955924137986]
  We propose a Vision Transformer Advanced by Exploring intrinsic IB from convolutions, i.e., ViTAE. ViTAE has several spatial pyramid reduction modules to downsample and embed the input image into tokens with rich multi-scale context. We obtain the state-of-the-art classification performance, i.e., 88.5% Top-1 classification accuracy on ImageNet validation set and the best 91.2% Top-1 accuracy on ImageNet real validation set.
  arXiv  Detail & Related papers  (2022-02-21T10:40:05Z)
• CSformer: Bridging Convolution and Transformer for Compressive Sensing [65.22377493627687]
  This paper proposes a hybrid framework that integrates the advantages of leveraging detailed spatial information from CNN and the global context provided by transformer for enhanced representation learning. The proposed approach is an end-to-end compressive image sensing method, composed of adaptive sampling and recovery. The experimental results demonstrate the effectiveness of the dedicated transformer-based architecture for compressive sensing.
  arXiv  Detail & Related papers  (2021-12-31T04:37:11Z)
• FeatMatch: Feature-Based Augmentation for Semi-Supervised Learning [64.32306537419498]
  We propose a novel learned feature-based refinement and augmentation method that produces a varied set of complex transformations. These transformations also use information from both within-class and across-class representations that we extract through clustering. We demonstrate that our method is comparable to current state of art for smaller datasets while being able to scale up to larger datasets.
  arXiv  Detail & Related papers  (2020-07-16T17:55:31Z)
• Group Equivariant Generative Adversarial Networks [7.734726150561089]
  In this work, we explicitly incorporate inductive symmetry priors into the network architectures via group-equivariant convolutional networks. Group-convariants have higher expressive power with fewer samples and lead to better gradient feedback between generator and discriminator.
  arXiv  Detail & Related papers  (2020-05-04T17:38:49Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.