VVS: Accelerating Visual Autoregressive Generation via Verification Skipping
Researchers propose VVS, a novel speculative decoding framework that accelerates visual autoregressive (AR) generation models by skipping verification steps. Visual AR models, despite strong image generation capabilities, suffer from high inference latency due to their next-token-prediction paradigm. Traditional speculative decoding uses a 'draft one step, then verify one step' approach, which does not reduce the number of forward passes. By exploiting the interchangeability of visual tokens, VVS explicitly cuts target model forward passes. The method is based on two observations: verification redundancy and stale feature reusability, which help maintain generation quality while improving speed. The work is published on arXiv with ID 2511.13587.
Key facts
- VVS is a speculative decoding framework for visual autoregressive generation.
- It reduces inference latency by skipping verification steps.
- Visual AR models use a next-token-prediction paradigm.
- Traditional speculative decoding does not reduce forward passes.
- VVS exploits interchangeability of visual tokens.
- Two key observations: verification redundancy and stale feature reusability.
- The paper is on arXiv with ID 2511.13587.
- The announcement type is replace-cross.
Entities
Institutions
- arXiv