Paper Overview
- Field: Computer Vision
- Authors: Shivam Duggal, Xingjian Bai, Zongze Wu, Richard Zhang, Eli Shechtman, Antonio Torralba, Phillip Isola, William T. Freeman
- Published: 2026-03-23
- arXiv: 2603.22283
- Tokenization infers latents from fully observed images.
- Generation infers latents from noise together with text or class conditioning.
- On both image and molecular modalities, UNITE achieves near state-of-the-art performance without adversarial losses or pretrained encoders (e.g., DINO).
- On ImageNet 256×256, the Base and Large models reach FID 2.12 and FID 1.73, respectively.
Abstract
Latent diffusion models (LDMs) enable high-fidelity synthesis by operating in learned latent spaces. However, training state-of-the-art LDMs requires complex staging: a tokenizer must be trained first, before the diffusion model can be trained in the frozen latent space.
The authors propose UNITE — an autoencoder architecture for unified tokenization and latent diffusion. UNITE consists of a Generative Encoder that serves as both image tokenizer and latent generator via weight sharing.
The key insight is that tokenization and generation can be viewed as the same latent inference problem under different conditioning regimes:
Method
Motivated by this insight, UNITE introduces a single-stage training procedure that jointly optimizes both tasks through two forward passes of the same generative encoder. Shared parameters allow gradients from both objectives to shape a common latent space, promoting a universal latent language.
Results
*Auto-collected on 2026-03-25.*