Fractional dynamics model speeds up diffusion transformers without extra training

The approach treats feature evolution as a non-Brownian diffusion process, enabling larger skip intervals during inference while preserving output quality.

Academic

Jinlong Yang · Lizilin · Jinke Wu · Yao Zhou

Sichuan University

Research Digest·10 November 2026·1 min read

Thread:Diffusion Inference Acceleration

Illustration: Fractional dynamics model speeds up diffusion transformers without extra training

In one line

FADE accelerates Diffusion Transformers by forecasting changing internal features with a fractional-dynamics model and a precomputed lookup table, maintaining quality during longer sampling skips.

Before you read this

An idea this paper builds on

diffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.70 papers →

Read it as

generate images and other media by repeatedly refining noise, which can make them slow to run. Skipping some steps saves time, but simply reusing old calculations can hurt the result because the model’s changes are uneven and sometimes abrupt.

FADE predicts those changing internal calculations, called features, with a forecasting rule designed for irregular patterns. It uses a lookup table, a stored set of values prepared in advance, to adapt predictions to different stages and parts of the model. It requires no retraining.

Tests across different models and media show that FADE can skip more steps while keeping generation quality strong, with negligible added delay. This could make high-quality generation more practical where computing time matters.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

diffusion transformer fractional anomalous dynamics feature caching training-free acceleration offline profiling