Scaling Diffusion Transformers Efficiently via $μ$P

TL;DR

Efficiently scaling diffusion Transformers via $μ$P, PixArt-$α$ surpasses baseline at 0.61B parameters.

cs.LG 🔴 Advanced 2025-05-21 39 views
Chenyu Zheng Xinyu Zhang Rongzhen Wang Wei Huang Zhi Tian Weilin Huang Jun Zhu Chongxuan Li
diffusion Transformers parametrization hyperparameter transfer generative models image generation

Key Findings

Methodology

The paper extends Maximal Update Parametrization ($μ$P) to diffusion Transformers, validating its effectiveness in models like U-ViT, DiT, PixArt-$α$, and MMDiT. Using Tensor Programs, it proves these models' $μ$P aligns with vanilla Transformers, enabling direct application of existing $μ$P methods.

Key Results

  • DiT-XL-2-$μ$P achieves 2.9x faster convergence with transferred learning rate compared to original DiT-XL-2.
  • PixArt-$α$ scales from 0.04B to 0.61B, with $μ$P models outperforming baselines at only 5.5% tuning cost.
  • MMDiT scales from 0.18B to 18B, with $μ$P models outperforming baselines at just 3% human expert cost.

Significance

$μ$P provides an efficient framework for scaling diffusion Transformers, significantly reducing hyperparameter tuning costs for large models. This supports further expansion of vision generative models, addressing high tuning cost issues.

Technical Contribution

The technical contribution lies in extending $μ$P theory from vanilla to diffusion Transformers, providing theoretical proof and practical validation, demonstrating $μ$P's hyperparameter transferability across widths, batch sizes, and training steps.

Novelty

This is the first application of $μ$P theory to diffusion Transformers, proving its applicability across different architectures and generative objectives, filling a gap in existing research.

Limitations

  • While $μ$P performs well in diffusion Transformers, its applicability to other generative models remains unverified.
  • Current research focuses on vision generation tasks, with other tasks' effectiveness yet to be explored.

Future Work

Future work could explore $μ$P's application in other generative models and its performance across different tasks and datasets. Further optimization of $μ$P's parametrization strategy could enhance model performance.

AI Executive Summary

Diffusion Transformers have become foundational in vision generative models but are limited by high hyperparameter tuning costs at large scales. This paper proposes applying Maximal Update Parametrization ($μ$P) to diffusion Transformers, validating its effectiveness in models like U-ViT, DiT, PixArt-$α$, and MMDiT.

Large-scale experiments demonstrate that $μ$P significantly enhances hyperparameter transferability, reducing tuning costs. For instance, DiT-XL-2-$μ$P achieves 2.9x faster convergence with transferred learning rate. PixArt-$α$, when scaled from 0.04B to 0.61B, sees $μ$P models outperform baselines at minimal tuning cost.

These findings establish $μ$P as an efficient framework for scaling diffusion Transformers, offering new research directions. However, $μ$P's applicability to other generative models needs further verification. Future work could explore $μ$P's performance across different tasks and datasets, optimizing its parametrization strategy for better model performance.

Deep Analysis

Background

Diffusion Transformers have become the backbone of modern vision generative models, applied in tasks like image and video generation. As datasets grow and task complexity increases, further scaling of diffusion Transformers is inevitable. However, the high cost of hyperparameter tuning for large models is a major bottleneck.

Core Problem

The core problem in scaling diffusion Transformers is the high cost of hyperparameter tuning, especially as model parameters reach billions. This limits the model's potential, highlighting the need for an efficient method to identify optimal hyperparameters.

Innovation

The core innovation is extending Maximal Update Parametrization ($μ$P) theory from vanilla to diffusion Transformers. Through theoretical proof and practical validation, it demonstrates $μ$P's hyperparameter transferability across different model widths, batch sizes, and training steps.

Methodology

  • �� Use Tensor Programs to prove diffusion Transformers' $μ$P aligns with vanilla Transformers.
  • �� Conduct systematic study of DiT-μP on ImageNet to verify hyperparameter transferability.
  • �� Apply $μ$P to text-to-image generation tasks with PixArt-$α$ and MMDiT, validating its effectiveness on large-scale models.

Experiments

Experiments include verifying DiT-μP's hyperparameter transferability on ImageNet and validating $μ$P's effectiveness in text-to-image generation tasks with PixArt-$α$ and MMDiT. Baselines include original DiT, PixArt-$α$, and MMDiT models.

Results

Results show DiT-XL-2-$μ$P achieves 2.9x faster convergence with transferred learning rate. PixArt-$α$, when scaled from 0.04B to 0.61B, sees $μ$P models outperform baselines at minimal tuning cost.

Applications

$μ$P can be used to scale large-scale vision generative models, particularly in scenarios requiring low tuning costs. Its successful application in text-to-image generation tasks suggests potential in multimodal generation tasks.

Limitations & Outlook

While $μ$P performs well in diffusion Transformers, its applicability to other generative models remains unverified. Current research focuses on vision generation tasks, with other tasks' effectiveness yet to be explored.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. Traditional methods require adjusting the seasoning every time you cook, much like traditional models needing extensive hyperparameter tuning. $μ$P is like a universal seasoning packet; you test it in a small pot and then apply it directly to a large pot. This way, you don't have to adjust the seasoning every time, saving time and effort.

ELI14 Explained like you're 14

Hey there! Imagine you're playing a game where every time you level up, you have to reset your character's gear and skills. Isn't that annoying? $μ$P is like a super power-up pack; you try it out at a low level and then use it directly at a high level, saving you a lot of hassle! Isn't that cool?

Glossary

Maximal Update Parametrization ($μ$P)

A parametrization method that allows stable hyperparameter transfer from small to large models, significantly reducing tuning costs.

Applied to diffusion Transformers to enhance scaling efficiency.

Diffusion Transformers

A model for vision generation that combines Transformer architecture with diffusion processes.

Serves as the backbone for modern vision generative models.

Hyperparameter Transfer

The process of directly applying optimal hyperparameters searched from small models to large models.

Achieved through $μ$P, reducing tuning costs for large models.

PixArt-$α$

A diffusion Transformer model for text-to-image generation.

Used in experiments to validate $μ$P's effectiveness.

MMDiT

A multimodal diffusion Transformer model combining image and text modalities.

Used in experiments to validate $μ$P's effectiveness.

Open Questions Unanswered questions from this research

  • 1 How can $μ$P be applied to other types of generative models? Current research focuses on vision generation tasks, leaving other tasks unexplored.
  • 2 Does $μ$P's parametrization strategy need adjustment to improve model performance across different tasks and datasets?

Applications

Immediate Applications

Scaling Vision Generative Models

Using $μ$P to reduce hyperparameter tuning costs for large-scale vision generative models, suitable for scenarios requiring rapid scaling.

Long-term Vision

Multimodal Generation Tasks

Applying $μ$P in multimodal generation tasks to enhance model scalability and performance, driving further development of generative models.

Abstract

Diffusion Transformers have emerged as the foundation for vision generative models, but their scalability is limited by the high cost of hyperparameter (HP) tuning at large scales. Recently, Maximal Update Parametrization ($μ$P) was proposed for vanilla Transformers, which enables stable HP transfer from small to large language models, and dramatically reduces tuning costs. However, it remains unclear whether $μ$P of vanilla Transformers extends to diffusion Transformers, which differ architecturally and objectively. In this work, we generalize standard $μ$P to diffusion Transformers and validate its effectiveness through large-scale experiments. First, we rigorously prove that $μ$P of mainstream diffusion Transformers, including U-ViT, DiT, PixArt-$α$, and MMDiT, aligns with that of the vanilla Transformer, enabling the direct application of existing $μ$P methodologies. Leveraging this result, we systematically demonstrate that DiT-$μ$P enjoys robust HP transferability. Notably, DiT-XL-2-$μ$P with transferred learning rate achieves 2.9 times faster convergence than the original DiT-XL-2. Finally, we validate the effectiveness of $μ$P on text-to-image generation by scaling PixArt-$α$ from 0.04B to 0.61B and MMDiT from 0.18B to 18B. In both cases, models under $μ$P outperform their respective baselines while requiring small tuning cost, only 5.5% of one training run for PixArt-$α$ and 3% of consumption by human experts for MMDiT-18B. These results establish $μ$P as a principled and efficient framework for scaling diffusion Transformers.

cs.LG cs.AI cs.CV