| name | stagevar-acceleration |
| title | StageVAR: Stage-Aware Acceleration for Visual Autoregressive Models |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.16483 |
| keywords | ["autoregressive","image-generation","acceleration","inference","optimization"] |
| description | Accelerate visual autoregressive (VAR) image generation 3.4× without retraining by analyzing generation stages. Exploits semantic irrelevance in detail-refinement stages where classifier-free guidance becomes redundant and features exhibit low-rank structure—enabling dimensionality reduction while preserving output quality. |
Overview
StageVAR addresses computational bottlenecks in visual autoregressive models by analyzing how image content is progressively established. Early stages build semantic structure, middle stages establish spatial arrangement, and late stages refine details. This stage structure reveals optimization opportunities unavailable in single-pass approaches.
Core Technique
The key insight is that generation stages have fundamentally different computational requirements.
Three-Stage Analysis Framework:
The method identifies distinct phases with different optimization potential:
class StageAwareVAR:
def analyze_generation(self, model):
"""
Identify three distinct generation stages with different
properties and optimization opportunities.
"""
stages = {
'semantic': {
'steps': 'early',
'property': 'establishes what image depicts',
'optimization': 'none (preserve)'
},
'structure': {
'steps': 'middle',
'property': 'defines spatial arrangement',
'optimization': 'none (preserve)'
},
'refinement': {
'steps': 'late',
'property': 'adds fine details',
'optimization': 'heavy (exploit low-rank, drop guidance)'
}
}
return stages
Semantic Irrelevance Exploitation:
In refinement stages, classifier-free guidance becomes unnecessary because text conditioning only affects high-level concepts, not fine details.
