Grouped Speculative Decoding for Autoregressive Image Generation

Junhyuk So, Juncheol Shin, Hyunho Kook, Eunhyeok Park · 2025

Read Grouped Speculative Decoding for Autoregressive Image Generation by Junhyuk So, Juncheol Shin, Hyunho Kook, Eunhyeok Park (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07747.txt.

Description

Junhyuk So and three co-authors speed up image-making models in a paper from 2025. These models build an image one token at a time, in sequence, which makes them slow. A known speedup drafts tokens ahead and checks them, but it accepts only the single most likely token. Image tokens, the authors show, often allow many valid choices at once.

Their method checks whole clusters of valid tokens instead of one token, and rebuilds the clusters on the fly. In tests it sped up image models by about 3.7 times on average with no loss of image quality. No extra training is needed, and the code is public.

Catalog record

CREATORJunhyuk So, Juncheol Shin, Hyunho Kook, Eunhyeok Park
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-07747

Browse the full library · About the Tower of Babel