Fixed-Point Distillation: One-Step Distillation of Discrete Diffusion Image Generators

Forum topic · 小凯 · 2026-05-22

Summary

This arXiv paper (2505.15984) by Chaoyang Wang and Yunhai Tong introduces Fixed-Point Distillation (FPD), an end-to-end framework for distilling discrete diffusion image generators into one-step models. Discrete diffusion models produce high-quality images but require slow iterative decoding. Prior single-step distillation approaches either train auxiliary score networks that double the compute or rely on specialized parameterizations and fragmented multi-stage pipelines. FPD instead constructs local correction targets by partially corrupting the student's one-step draft and refining it with a single teacher step. Discrete tokens are lifted into continuous feature space, where a multi-bandwidth drift loss iteratively accumulates corrections. A straight-through estimator feeds exact hard-sampled tokens to the teacher and decoder in the forward pass, ensuring training and inference operate on the same codebook manifold while routing continuous gradients back to the student logits. The fully differentiable path also supports an optional unconditional adversarial objective for perceptual realism. Evaluations on class-conditional and text-to-image generation show FPD achieves competitive visual fidelity and structural alignment in a single step, narrowing the gap to multi-step teachers and outperforming existing discrete distillation baselines.

Paper Overview

Field: Computer Vision Authors: Chaoyang Wang, Yunhai Tong Released: 2025-05-20 arXiv: 2505.15984

Introduction

Discrete diffusion models excel at visual synthesis but rely on slow, iterative decoding. Existing single-step distillation methods attempt to bypass this bottleneck, either by training auxiliary score networks that effectively double compute, or by introducing specialized parameterizations and multi-stage pipelines that fragment optimization.

Method: Fixed-Point Distillation (FPD)

This paper introduces Fixed-Point Distillation (FPD), an end-to-end framework with the following key ideas:

Local correction targets: FPD constructs local correction targets by partially corrupting the student's one-step draft and refining it with a single teacher step.
Multi-bandwidth drift loss: Discrete tokens are lifted into continuous feature space, where a multi-bandwidth drift loss iteratively accumulates these corrections in a semantically meaningful space.
Straight-through estimator: To backpropagate through the discrete bottleneck, a straight-through estimator feeds exact hard-sampled tokens to the teacher and decoder in the forward pass. This ensures training and inference operate on the same codebook manifold while routing continuous gradients back to the student logits.
Optional adversarial objective: The fully differentiable path also accommodates an optional unconditional adversarial objective to enhance perceptual realism.

Results

Evaluations on class-conditional and text-conditional generation validate the framework's effectiveness. FPD achieves competitive visual fidelity and structural alignment with single-step inference, narrowing the gap to multi-step teachers while outperforming existing discrete distillation baselines.