DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision

TL;DR

DiffRatio introduces a novel method for training one-step diffusion models by directly estimating score differences via learned density ratios, reducing bias and improving efficiency and quality over traditional teacher-supervised distillation methods.

Contribution

The paper proposes a new framework, DiffRatio, that simplifies training, reduces bias, and enhances efficiency by directly estimating score differences with a lightweight density-ratio network.

Findings

Outperforms most teacher-supervised methods on CIFAR-10 and ImageNet.

Reduces gradient estimation bias and auxiliary network size.

Enables inference-time scaling for improved quality without extra training.

Abstract

Score-based distillation methods (e.g., variational score distillation) train one-step diffusion models by first pre-training a teacher score model and then distilling it into a one-step student model. However, the gradient estimator in the distillation stage usually suffers from two sources of bias: (1) biased teacher supervision due to score estimation error incurred during pre-training, and (2) the student model's score estimation error during distillation. These biases can degrade the quality of the resulting one-step diffusion model. To address this, we propose DiffRatio, a new framework for training one-step diffusion models: instead of estimating the teacher and student scores independently and then taking their difference, we directly estimate the score difference as the gradient of a learned log density ratio between the student and data distributions across diffusion time steps. This approach greatly simplifies the training pipeline, significantly reduces gradient estimation bias, and improves one-step generation quality. Additionally, it also reduces auxiliary network size by using a lightweight density-ratio network instead of two full score networks, which improves computational and memory efficiency. DiffRatio achieves competitive one-step generation results on CIFAR-10 and ImageNet (64x64 and 512x512), outperforming most teacher-supervised distillation methods. Moreover, the learned density ratio naturally serves as a verifier, enabling a principled inference-time parallel scaling scheme that further improves the generation quality without external rewards or additional sequential computation.