What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion

Researchers introduced Prior-Aligned AutoEncoder (PAE), an open-source latent diffusion tokenizer designed to better match the needs of diffusion models.

PAE aligns latent geometry with the prior, improving generation quality and enabling much faster downstream DiT training convergence.

On ImageNet 256×256, the method reports strong few-step sampling and state-of-the-art generation performance.

The approach suggests latent diffusion models can become both easier to train and more sample-efficient without hurting reconstruction quality.