Scaling Up Parameter Generation: A Recurrent Diffusion Approach

Abstract

Parameter generation has long struggled to match the scale of today’s large vision and language models, curbing its broader utility. In this paper, we introduce Recurrent Diffusion for Large-Scale Parameter Generation (RPG), a novel framework that generates full neural network parameters—up to hundreds of millions—on a single GPU. Our approach first partitions a network’s parameters into non-overlapping ‘tokens’, each corresponding to a distinct portion of the model. A recurrent mechanism then learns the inter-token relationships, producing ‘prototypes’ which serve as conditions for a diffusion process that ultimately synthesizes the full parameters. Across a spectrum of architectures and tasks—including ResNets, ConvNeXts and ViTs on ImageNet-1K and COCO, and even LoRA-based LLMs—RPG achieves performance on par with fully trained networks while avoiding excessive memory overhead. Notably, it generalizes beyond its training set to generate valid parameters for previously unseen tasks, highlighting its flexibility in dynamic and open-ended scenarios. By overcoming the longstanding memory and scalability barriers, RPG serves as a critical advance in ‘AI generating AI’, potentially enabling efficient weight generation at scales previously deemed infeasible.

Environment

Before you get started, you need to set up a conda environment first.

Create your conda environment.

conda create -n rpg python=3.11
conda activate rpg
conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

Install mamba-ssm. (You may run into compilation issues, refer to the official mamba-ssm repository for details.)

pip install causal-conv1d
pip install mamba-ssm[causal-conv1d]

Install other dependencies for this repository.

git lfs install
git clone https://huggingface.co/MTDoven/Recurrent-Parameter-Generation.git
cd Recurrent-Parameter-Generation
pip install -r requirements.txt

Quick Start

Try to generate with RPG model.

cd ./workspace
CUDA_VISIBLE_DEVICES=0 sh demo.sh
# CUDA_VISIBLE_DEVICES=<GPU_index> sh demo.sh

Here are some examples.

description: "Give me a model to select all living things"
expected_class: [0,0,1,1,1,1,1,1,0,0]  # bird, cat, deer, dog, frog, horse

description: "Find all vehicles that operate on roads"
expected_class: [0,1,0,0,0,0,0,0,0,1]  # automobile, truck

description: "Select all things that can fly"
expected_class: [1,0,1,0,0,0,0,0,0,0]  # airplane, bird

description: "Find all transportation methods that travel on water"
expected_class: [0,0,0,0,0,0,0,0,1,0]  # ship

description: "Classify all mammals"
expected_class: [0,0,0,1,1,1,0,1,0,0]  # cat, deer, dog, horse

description: "Find all animals with fur"
expected_class: [0,0,1,1,1,1,0,1,0,0]  # bird, cat, deer, dog, horse

description: "Select all pets commonly found in households"
expected_class: [0,0,1,1,0,1,0,0,0,0]  # bird, cat, dog

description: "Identify all cold-blooded animals"
expected_class: [0,0,0,0,0,0,1,0,0,0]  # frog

description: "Find all objects that can carry cargo"
expected_class: [1,1,0,0,0,0,0,0,1,1]  # airplane, automobile, ship, truck

description: "Select all things used for commercial transportation"
expected_class: [1,1,0,0,0,0,0,0,1,1]  # airplane, automobile, ship, truck

description: "Identify all animals that can swim naturally"
expected_class: [0,0,0,1,0,0,1,0,0,0]  # cat, frog

description: "Find all things with wheels"
expected_class: [1,1,0,0,0,0,0,0,0,1]  # airplane, automobile, truck

description: "Select all creatures with four legs"
expected_class: [0,0,0,1,1,1,0,1,0,0]  # cat, deer, dog, horse

description: "Identify all creatures that live in forests"
expected_class: [0,0,1,1,1,1,0,0,0,0]  # bird, cat, deer, dog

description: "Find all animals that can live near water"
expected_class: [0,0,1,0,0,0,1,0,0,0]  # bird, frog

description: "Select all man-made objects"
expected_class: [1,1,0,0,0,0,0,0,1,1]  # airplane, automobile, ship, truck

description: "Find all things that make noise naturally"
expected_class: [0,0,1,1,1,1,1,1,0,0]  # all animals

description: "Identify all animals that can climb trees"
expected_class: [0,0,1,1,0,1,0,0,0,0]  # bird, cat, dog

"Select all animals that hunt other animals"
expected_class: [0,0,0,1,0,1,0,0,0,0]  # cat, dog

description: "Find all things that are both man-made and can operate on water"
expected_class: [0,0,0,0,0,0,0,0,1,0]  # ship

description: "Select all animals that are both pets and can climb"
expected_class: [0,0,0,1,0,1,0,0,0,0]  # cat, dog

You can get more information from Github and Project-Page.

Acknowledgment

We thank Zhiyuan Liang, Zhuang Liu, Gongfan Fang, Xuanlei Zhao, Yuhao Zhou, Mingjia Shi, Zangwei Zheng, Ziheng Qin, Tianlong Chen, and Zhangyang Wang for valuable discussions and feedbacks. This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-008).

Citation

@misc{wang2025recurrent,
      title={Recurrent Diffusion for Large-Scale Parameter Generation},
      author={Wang, Kai and Tang, Dongwen and Zhao, Wangbo and You, Yang},
      year={2025},
}

MTDoven
/

Recurrent-Parameter-Generation