Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders

TL;DR

This paper demonstrates that Representation Autoencoders (RAEs) can be scaled effectively for large-scale text-to-image diffusion models, outperforming VAEs in fidelity, stability, and convergence speed, and enabling unified multimodal reasoning.

Contribution

The work shows that scaling RAEs simplifies the diffusion framework and improves performance over VAEs for large-scale text-to-image generation, with better stability and convergence.

Findings

RAEs outperform VAEs across all model scales during pretraining.

RAEs remain stable during extended finetuning, unlike VAEs which overfit.

RAEs achieve faster convergence and higher quality in image generation.

Abstract

Representation Autoencoders (RAEs) have shown distinct advantages in diffusion modeling on ImageNet by training in high-dimensional semantic latent spaces. In this work, we investigate whether this framework can scale to large-scale, freeform text-to-image (T2I) generation. We first scale RAE decoders on the frozen representation encoder (SigLIP-2) beyond ImageNet by training on web, synthetic, and text-rendering data, finding that while scale improves general fidelity, targeted data composition is essential for specific domains like text. We then rigorously stress-test the RAE design choices originally proposed for ImageNet. Our analysis reveals that scaling simplifies the framework: while dimension-dependent noise scheduling remains critical, architectural complexities such as wide diffusion heads and noise-augmented decoding offer negligible benefits at scale Building on this simplified framework, we conduct a controlled comparison of RAE against the state-of-the-art FLUX VAE across diffusion transformer scales from 0.5B to 9.8B parameters. RAEs consistently outperform VAEs during pretraining across all model scales. Further, during finetuning on high-quality datasets, VAE-based models catastrophically overfit after 64 epochs, while RAE models remain stable through 256 epochs and achieve consistently better performance. Across all experiments, RAE-based diffusion models demonstrate faster convergence and better generation quality, establishing RAEs as a simpler and stronger foundation than VAEs for large-scale T2I generation. Additionally, because both visual understanding and generation can operate in a shared representation space, the multimodal model can directly reason over generated latents, opening new possibilities for unified models.