FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation

TL;DR

FlowLM is a novel language model that transforms pre-trained diffusion models into efficient flow-based models, enabling high-quality few-step text generation with minimal training.

Contribution

We introduce FlowLM, a method to convert diffusion language models into flow models via fine-tuning, achieving superior few-step generation performance with fewer training epochs.

Findings

FlowLM rivals 2000-step diffusion sampling quality with fewer steps.

Fine-tuning FlowLM saturates performance in half the epochs of training from scratch.

Using clean data prediction improves the flow matching training objective.

Abstract

We present FlowLM, a flow matching language model transformed from pre-trained diffusion language models via efficient fine-tuning. By re-aligning the curved sampling trajectories of diffusion models into straight-line flows, FlowLM enables high quality few-step generation that rivals or even outperforms the quality of 2,000-step diffusion sampling with very few training epochs. Remarkably, finetuned FlowLM reaches performance saturation with only half as many training epochs as training from scratch, both approaches greatly outperforming the original diffusion model, thereby validating our method. Furthermore, we validate a more effective training objective for flow matching: predicting clean data to consistently guide the sampling process towards the true data distribution. Empirical results demonstrate that our approach is highly effective for high-quality, few-step text generation.