Researchers at New York University have developed a new architecture for diffusion models that improves the semantic representation of the images they generate. “Diffusion Transformer with Representation Autoencoders” (RAE) challenges some of the accepted norms of building diffusion models. The NYU researcher's model is more efficient and accurate than standard diffusion models, takes advantage of the latest research in representation learning and could pave the way for new applications that were previously too difficult or expensive.
This breakthrough could unlock more reliable and powerful features for enterprise applications. "To edit images well, a model has to really understand what’s in them," paper co-author Saining Xie told VentureBeat. "RAE helps connect that understanding part with the generation part." He also pointed to future applications in "RAG-based generation, where you use RAE encoder features for search and then generate new images based on the search results," as well as in "video generation and action-conditioned world models."
The state of generative modeling
Diffusion models, the technology behind most of today’s powerful image generators, frame generation as a process of learning to compress and decompress images. A variational autoencoder (VAE) learns a compact representation of an image’s key features in a so-called “latent space.” The model is then trained to generate new images by reversing this process from random noise.
While the diffusion part of these models has advanced, the autoencoder used in most of them has remained largely unchanged in recent years. According to the NYU researchers, this standard autoencoder (SD-VAE) is suitable for capturing low-level features and local appearance, but lacks the “global semantic structure crucial for generalization and generative performance.”
At the same time, the field has seen impressive advances in image representation learning with models such as DINO, MAE and CLIP. These models learn semantically-structured visual features that generalize across tasks and can serve as a natural basis for visual understanding. However, a widely-held belief has kept devs from using these architectures in image generation: Models focused on semantics are not suitable for generating images because they don’t capture granular, pixel-level features. Practitioners also believe that diffusion models do not work well with the kind of high-dimensional representations that semantic models produce.
Diffusion with representation encoders
The NYU researchers propose replacing the standard VAE with “representation autoencoders” (RAE). This new type of autoencoder pairs a pretrained representation encoder, like Meta’s DINO, with a trained vision transformer decoder. This approach simplifies the training process by using existing, powerful encoders that have already been trained on massive datasets.
To make this work, the team developed a variant of the diffusion transformer (DiT), the backbone of most image generation models. This modified DiT can be trained efficiently in the high-dimensional space of RAEs without incurring huge compute costs. The researchers show that frozen representation encoders, even those optimized for semantics, can be adapted for image generation tasks. Their method yields reconstructions that are superior to the standard SD-VAE without adding architectural complexity.
However, adopting this approach requires a shift in thinking. "RAE isn’t a simple plug-and-play autoencoder; the diffusion modeling part also needs to evolve," Xie explained. "One key point we want to highlight is that latent space modeling and generative modeling should be co-designed rather than treated separately."
With the right architectural adjustments, the researchers found that higher-dimensional representations are an advantage, offering richer structure, faster convergence and better generation quality. In their paper, the researchers note that these "higher-dimensional latents introduce effectively no extra compute or memory costs." Furthermore, the standard SD-VAE is more computationally expensive, requiring about six times more compute for the encoder and three times more for the decoder, compared to RAE.
Stronger performance and efficiency
The new model architecture delivers significant gains in both training efficiency and generation quality. The team's improved diffusion recipe achieves strong results after only 80 training epochs. Compared to prior diffusion models trained on VAEs, the RAE-based model achieves a 47x training speedup. It also outperforms recent methods based on representation alignment with a 16x training speedup. This level of efficiency translates directly into lower training costs and faster model development cycles.
For enterprise use, this translates into more reliable and consistent outputs. Xie noted that RAE-based models are less prone to semantic errors seen in classic diffusion, adding that RAE gives the model "a much smarter lens on the data." He observed that leading models like ChatGPT-4o and Google's Nano Banana are moving toward "subject-driven, highly consistent and knowledge-augmented generation," and that RAE's semantically rich foundation is key to achieving this reliability at scale and in open source models.
The researchers demonstrated this performance on the ImageNet benchmark. Using the Fréchet Inception Distance (FID) metric, where a lower score indicates higher-quality images, the RAE-based model achieved a state-of-the-art score of 1.51 without guidance. With AutoGuidance, a technique that uses a smaller model to steer the generation process, the FID score dropped to an even more impressive 1.13 for both 256×256 and 512×512 images.
By successfully integrating modern representation learning into the diffusion framework, this work opens a new path for building more capable and cost-effective generative models. This unification points toward a future of more integrated AI systems.
"We believe that in the future, there will be a single, unified representation model that captures the rich, underlying structure of reality… capable of decoding into many different output modalities," Xie said. He added that RAE offers a unique path toward this goal: "The high-dimensional latent space should be learned separately to provide a strong prior that can then be decoded into various modalities — rather than relying on a brute-force approach of mixing all data and training with multiple objectives at once."
