Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

I-Max: Maximize the Resolution Potential of Pre-trained Rectified Flow Transformers with Projected Flow

Ruoyi Du, Dongyang Liu, Le Zhuo, Qin Qi, Hongsheng Li, Zhanyu Ma, Peng Gao

TL;DR

I-Max tackles the problem of tuning-free, high-resolution extrapolation for pre-trained Rectified Flow Transformers by introducing Projected Flow, which leverages the low-dimensional projections of high-resolution flows to stabilize extrapolation, and an Inference Toolkit that adjusts RoPE, SNR, entropy, and token-ratio dynamics for better generalization. The approach enables stable extrapolation up to about $4\times$ to $16\times$ beyond native resolutions, reaching $4096^2$ on models like Lumina-Next-2K and Flux.1-dev, with visible improvements in local detail and artifact correction while maintaining practical inference stability. This work provides a practical path to higher-resolution diffusion outputs without additional tuning, expanding the usable resolution range of pre-trained Text-to-Image Rectified Flow Transformers.

Abstract

Rectified Flow Transformers (RFTs) offer superior training and inference efficiency, making them likely the most viable direction for scaling up diffusion models. However, progress in generation resolution has been relatively slow due to data quality and training costs. Tuning-free resolution extrapolation presents an alternative, but current methods often reduce generative stability, limiting practical application. In this paper, we review existing resolution extrapolation methods and introduce the I-Max framework to maximize the resolution potential of Text-to-Image RFTs. I-Max features: (i) a novel Projected Flow strategy for stable extrapolation and (ii) an advanced inference toolkit for generalizing model knowledge to higher resolutions. Experiments with Lumina-Next-2K and Flux.1-dev demonstrate I-Max's ability to enhance stability in resolution extrapolation and show that it can bring image detail emergence and artifact correction, confirming the practical value of tuning-free resolution extrapolation.

I-Max: Maximize the Resolution Potential of Pre-trained Rectified Flow Transformers with Projected Flow

TL;DR

I-Max tackles the problem of tuning-free, high-resolution extrapolation for pre-trained Rectified Flow Transformers by introducing Projected Flow, which leverages the low-dimensional projections of high-resolution flows to stabilize extrapolation, and an Inference Toolkit that adjusts RoPE, SNR, entropy, and token-ratio dynamics for better generalization. The approach enables stable extrapolation up to about to beyond native resolutions, reaching on models like Lumina-Next-2K and Flux.1-dev, with visible improvements in local detail and artifact correction while maintaining practical inference stability. This work provides a practical path to higher-resolution diffusion outputs without additional tuning, expanding the usable resolution range of pre-trained Text-to-Image Rectified Flow Transformers.

Abstract

Rectified Flow Transformers (RFTs) offer superior training and inference efficiency, making them likely the most viable direction for scaling up diffusion models. However, progress in generation resolution has been relatively slow due to data quality and training costs. Tuning-free resolution extrapolation presents an alternative, but current methods often reduce generative stability, limiting practical application. In this paper, we review existing resolution extrapolation methods and introduce the I-Max framework to maximize the resolution potential of Text-to-Image RFTs. I-Max features: (i) a novel Projected Flow strategy for stable extrapolation and (ii) an advanced inference toolkit for generalizing model knowledge to higher resolutions. Experiments with Lumina-Next-2K and Flux.1-dev demonstrate I-Max's ability to enhance stability in resolution extrapolation and show that it can bring image detail emergence and artifact correction, confirming the practical value of tuning-free resolution extrapolation.

Paper Structure

This paper contains 24 sections, 5 equations, 15 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Preliminary: Rectified Flow
  4. Projected Flow
  5. Inference Toolkit
  6. Denoise beyond the training resolution.
  7. Balance the SNR shift.
  8. Balance the entropy shift of self-attention.
  9. Balance the image/text sequence length ratio for MMDiT.
  10. Experiments
  11. Implementation Details
  12. Effectiveness of Each Component of I-Max
  13. Gains from Resolution Extrapolation
  14. Comparison of Different Low-resolution Guidance
  15. Why Do We Need Low-resolution Guidance?
  16. ...and 9 more sections

Figures (15)

  • Figure 1: Landscape images crafted by Lumina-Next-2K and Flux.1-dev equipped with I-Max.
  • Figure 2: Extrapolation of 1K$\to$4K on Flux.1-dev enhances the generated result with richer local details and corrects artifacts in small objects.
  • Figure 3: Illustration of Projected Flow Mechanism:(a) In the low-dimensional space of native resolution, RFTs accurately predict the flow direction, ensuring precise distribution transfer. (b) In the high-dimensional space of extrapolated resolution, RFTs struggle to accurately predict the flow direction, degrading the quality of distribution transfer. (c) Projected Flow treats the flow in the low-dimensional space as a deterministic projection of the flow in the high-dimensional space, reducing the difficulty of predicting the flow direction at extrapolated resolutions.
  • Figure 4: Sequential ablation of I-Max. We illustrate the effect of sequentially removing different components of I-Max on Flux.1-dev across the 1K$\to$4K resolution range.
  • Figure 5: Single ablation results of I-Max.
  • ...and 10 more figures