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

Forum topic · 小凯 · 2026-05-22

Summary

Discrete diffusion models excel at visual synthesis but depend on slow iterative decoding. This paper introduces Fixed-Point Distillation (FPD), an end-to-end framework for distilling discrete diffusion image generators into efficient one-step generators. FPD 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 in a semantically meaningful space. Training uses a straight-through estimator that feeds exact hard-sampled tokens to the teacher and decoder, ensuring training and inference operate on the same codebook manifold while routing continuous gradients back to the student logits. This fully differentiable path also supports an optional unconditional adversarial objective for improved perceptual realism. Evaluations on class-conditional and text-to-image generation show that FPD achieves competitive visual fidelity and structural alignment with single-step inference, narrowing the gap to multi-step teachers and outperforming existing discrete diffusion distillation baselines. Paper: arXiv 2505.15984 by Chaoyang Wang and Yunhai Tong, released May 2025.

Paper Overview

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

Background

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)

FPD is an end-to-end framework that constructs local correction targets by partially corrupting the student's one-step draft and refining it with a single teacher step.

Key components:

Multi-bandwidth drift loss: Discrete tokens are lifted into continuous features, and a multi-bandwidth drift loss iteratively accumulates corrections, computing the training objective in a semantically meaningful space.
Straight-through estimator: To backpropagate through the discrete bottleneck, exact hard-sampled tokens are fed to the teacher and decoder in the forward pass, ensuring training and inference operate on the same codebook manifold while continuous gradients are routed back to the student logits.
Adversarial objective: This fully differentiable path additionally accommodates an optional unconditional adversarial objective to enhance perceptual realism.

Results

Evaluations on class-conditional and text-conditioned 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 surpassing existing discrete distillation baselines.

---

*Auto-collected on 2026-05-22*