WeightFormer

Linear-Time Global Visual Modeling without Explicit Attention

arXiv

Ruize He*, Dongchen Han*, Gao Huang ✉️

Tsinghua University

Existing research largely attributes the global sequence modeling capability of Transformers to the explicit computation of attention weights, a process that inherently incurs quadratic computational complexity. In this work, we offer a novel perspective: we demonstrate that self-attention can be mathematically reframed as a Multi-Layer Perceptron (MLP) equipped with dynamically predicted parameters. Through this lens, we explain attention's global modeling power not as explicit token-wise aggregation, but as an implicit process where dynamically generated parameters act as a compressed representation of the global context. Inspired by this insight, we investigate a fundamental question: can we achieve Transformer-level sequence global modeling entirely through dynamic parameterization while maintaining linear complexity, effectively replacing explicit attention? To explore this, we design various dynamic parameter prediction strategies and integrate them into standard network layers. Extensive empirical studies on vision models demonstrate that dynamic parameterization can indeed serve as a highly effective, linear-complexity alternative to explicit attention, opening new pathways for efficient sequence modeling.

Overview

Installation

Option 1: uv (recommended)

uv sync
source .venv/bin/activate

Option 2: Conda

conda create -n weightformer python=3.12 -y
conda activate weightformer
pip install -r requirements.txt

Dataset Preparation

Prepare ImageNet-1K in the standard format:

imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class2
    │   ├── img2.jpeg
    │   └── ...
    └── ...

Then update the dataset path in:

cfg/*.yaml

Image Classification

Training from scratch

torchrun --nproc_per_node=8 main.py --cfg cfg/wfm_t.yaml

Evaluation

torchrun --nproc_per_node=8 main.py --eval --cfg cfg/wfm_t.yaml --resume wfm-t.pth

Replace wfm_t.yaml with your desired config for the T, S, or B variants.

Image Generation

To use WeightFormer in DiT-style image generation, replace DiT's models.py with model/wfm_dit.py.

Follow the setup instructions from fast-DiT (recommended) or DiT for training and evaluation.

Weights and Logs

Acknowledgements

This project is built upon DeiT, Swin Transformer, and DiT.

Citation

@article{he2024weightformer,
  title={Linear-Time Global Visual Modeling without Explicit Attention},
  author={He, Ruize and Han, Dongchen and Huang, Gao},
  journal={arXiv preprint arXiv:2605.01711},
  year={2026}
}