Speculative Forcing: Efficient Autoregressive Video Generation with Early Token Acceptance

Forum topic · 小凯 · 2026-10-17

Summary

Speculative Forcing is a framework for accelerating few-step autoregressive video generators, which currently run every token through the final denoising step regardless of how much refinement each token actually needs. The authors observe that many tokens already closely match their eventual full-step outputs in the initial steps, so uniform computation is often unnecessary. The method treats intermediate clean estimates as drafts and predicts when individual tokens can be safely accepted early. Once accepted, a token is frozen as its final output with no further denoising, while its cached keys and values remain available as context for remaining tokens. Two verifier designs are explored, including a training-free one based on similarity to clean estimates. On LongLive, the method matches a four-step generator using roughly 1.7 denoising steps per token, running over 1.5x faster and reaching 41.7 and 37.8 fps. On Lingbot-World 2.0, it maintains quality with about 2.5 denoising steps while running 1.35x faster. Paper: arXiv 2610.22347.

Overview

Field: Computer Vision (CV)
Authors: Hyunwook Choi, Seungryong Kim, Dongyoon Han, Sangdoo Yun
Published: 2026-10-15
arXiv: 2610.22347

Key Idea

Few-step autoregressive video generators enable streaming generation, yet they run every token through the final denoising step, regardless of how much refinement it needs. The authors find this uniform computation is often unnecessary: many tokens already closely match their eventual full-step outputs in the initial steps, while others require refinement until the end.

Method

Speculative Forcing treats intermediate clean estimates as drafts and predicts when individual tokens can be safely accepted early.

Once accepted, a token is frozen as its final output and undergoes no further denoising.
Its cached keys and values remain available as context for the remaining tokens.
Two alternative verifier designs are explored, including a training-free one based on similarity to clean estimates.

Results

LongLive: matches a four-step generator with ~1.7 denoising steps per token, runs over 1.5x faster, reaching 41.7 and 37.8 fps.
Lingbot-World 2.0: maintains quality with ~2.5 denoising steps while running 1.35x faster.