CIAR: Interval-based Collaborative Decoding for Image Generation Acceleration

2026-03-26

AI Summary

This work addresses the computational inefficiency of autoregressive image generation models, whose sequential decoding hinders efficient on-device deployment. To overcome this limitation, the authors propose CIAR, a cloud-edge collaborative framework that introduces continuous probability intervals—replacing discrete token sets—to enable on-device quantification and skipping of visually redundant regions. Through interval-augmented decoding and distribution alignment training, CIAR achieves significant acceleration in generation while preserving image fidelity and semantic consistency. Compared to existing approaches, it reduces inference latency by 2.18× and decreases cloud requests by 70%. Key innovations include the continuous probability interval representation, a token-level redundancy-aware mechanism on the device, and a novel collaborative inference architecture.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Learning on the Edge & Model Compression
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Auto-regressive (AR) models have recently made notable progress in image generation, achieving performance comparable to diffusion-based approaches. However, their computational intensity and sequential nature impede on-device deployment, causing disruptive latency. We address this via a cloud-device collaboration framework \textbf{CIAR}, which utilizes on-device self-verification to handle two key properties of visual synthesis: \textit{the vast token vocabulary} required for high-fidelity images and \textit{inherent spatial redundancy} which leads to extreme predictability in homogeneous regions, while object boundaries exhibit high uncertainty. Uniform verification wastes resources on such redundant tokens. Our solution centers on an on-device token uncertainty quantifier, which adopts continuous probability intervals to accelerate processing and make it feasible for large visual vocabularies instead of conventional discrete solution sets. Additionally, we incorporate a Interval-enhanced decoding module to further speed up decoding while maintaining visual fidelity and semantic consistency via a distribution alignment training strategy. Extensive experiments demonstrate that CIAR achieves a 2.18x speed-up and reduces cloud requests by 70\%, while preserving image quality compared to existing methods.

Problem

• auto-regressive image generation
• on-device deployment
• computational latency
• token redundancy
• visual synthesis

Innovation

• interval-based decoding
• collaborative inference
• token uncertainty quantification
• visual vocabulary acceleration
• distribution alignment