Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation

TL;DR

Smoothie introduces a novel diffusion method that smooths token embeddings based on semantic similarity, improving text generation quality by combining continuous and discrete diffusion advantages.

Contribution

It proposes a new diffusion approach on token embeddings that enhances text generation by balancing semantic structure and discreteness.

Findings

Outperforms existing diffusion models in generation quality.

Smoothing diffusion on token embeddings yields better performance than standard embedding or categorical simplex.

Code is publicly available at https://github.com/ashaba1in/smoothie.

Abstract

Diffusion models have achieved state-of-the-art performance in generating images, audio, and video, but their adaptation to text remains challenging due to its discrete nature. Prior approaches either apply Gaussian diffusion in continuous latent spaces, which inherits semantic structure but struggles with token decoding, or operate in categorical simplex space, which respect discreteness but disregard semantic relation between tokens. In this paper, we propose Smoothing Diffusion on Token Embeddings (Smoothie), a novel diffusion method that combines the strengths of both approaches by progressively smoothing token embeddings based on semantic similarity. This technique enables gradual information removal while maintaining a natural decoding process. Experimental results on several sequence-to-sequence and unconditional generation tasks demonstrate that Smoothie outperforms existing diffusion-based models in generation quality. Furthermore, ablation studies show that our proposed diffusion space yields better performance than both the standard embedding space and the categorical simplex. The code is available at https://github.com/ashaba1in/smoothie.