Greedy Growing Enables High-Resolution Pixel-Based Diffusion Models

TL;DR

This paper introduces a simple greedy growing method for training high-resolution pixel-based diffusion models, eliminating the need for cascaded super-resolution components and enabling stable, large-scale image generation.

Contribution

The authors propose a novel greedy architecture growth algorithm that stabilizes training and scales diffusion models to high resolutions without cascades or additional regularization.

Findings

Able to train models up to 8B parameters without extra regularization

Achieved high-resolution 1024x1024 image generation with superior human preference

Eliminated the need for cascaded super-resolution in diffusion models

Abstract

We address the long-standing problem of how to learn effective pixel-based image diffusion models at scale, introducing a remarkably simple greedy growing method for stable training of large-scale, high-resolution models. without the needs for cascaded super-resolution components. The key insight stems from careful pre-training of core components, namely, those responsible for text-to-image alignment {\it vs.} high-resolution rendering. We first demonstrate the benefits of scaling a {\it Shallow UNet}, with no down(up)-sampling enc(dec)oder. Scaling its deep core layers is shown to improve alignment, object structure, and composition. Building on this core model, we propose a greedy algorithm that grows the architecture into high-resolution end-to-end models, while preserving the integrity of the pre-trained representation, stabilizing training, and reducing the need for large high-resolution datasets. This enables a single stage model capable of generating high-resolution images without the need of a super-resolution cascade. Our key results rely on public datasets and show that we are able to train non-cascaded models up to 8B parameters with no further regularization schemes. Vermeer, our full pipeline model trained with internal datasets to produce 1024x1024 images, without cascades, is preferred by 44.0% vs. 21.4% human evaluators over SDXL.