Multi-Scale Local Speculative Decoding for Image Generation

Elia PeruzzoGuillaume SautièreAmirhossein Habibian · 2026

Read Multi-Scale Local Speculative Decoding for Image Generation by Elia PeruzzoGuillaume SautièreAmirhossein Habibian (2026) 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-2601-05149.txt.

Description

Elia Peruzzo, Guillaume Sautière, and Amirhossein Habibian speed up image builds that work one bit at a time in a 2026 paper. Such builds have a top rate for how true their shots look, but their step-by-step way is slow by birth. A known speedup has a small model draft bits for a big model to check. For shots, though, a draft bit is vague, and the check knows naught of where bits sit in the shot.

Their frame, MuLo-SD, drafts at low sharp grade, then blows the draft up for the check. That weds multi-grade drafts to checks that know the lay of the shot. More drafts should pass the check, and that pass rate is where the speed comes from. Fast shot builds cut the cost of art tools that lean on such step-wise models.

Catalog record

CREATORElia PeruzzoGuillaume SautièreAmirhossein Habibian
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-05149

Browse the full library · About the Tower of Babel