Improved Baselines with Representation Autoencoders

Researchers revised representation autoencoder design for vision diffusion models and found that simpler choices can work better.

Summed encoder layers, representation alignment, and output reparameterization improved reconstruction, guidance, and training stability.

The resulting RAEv2 model achieved state-of-the-art results on ImageNet-256 and other tasks with far fewer training epochs.

The work suggests a more efficient path for training image generation models with stronger quality and faster convergence.