HybridStitch：面向高效扩散生成的 Token 自适应模型拼接

HybridStitch: Token-Adaptive Model Stitching for Efficient Diffusion Generation

AI 导读

HybridStitch 是一个免训练框架，通过 token 级反馈为每个更新选择提供模型的来源，在四个主要基准上实现图像生成提速 1.63–1.68 倍、扩散大语言模型（dLLM）吞吐量提升 1.15–1.58 倍。该方法让连续图像精修与离散语言承诺共享同一生成状态，并以模型专属 KV 缓冲为选定查询提供注意力上下文。dLLM 解码查询成本估计降低 50–66%，生成质量相当。

正文 · 原文

View PDF HTML (experimental)

Abstract:Iterative diffusion generation requires different amounts of model capacity at different token positions. Prior model combinations allocate capacity by denoising step or by drafting and verification role. We present HybridStitch, a training-free framework that uses token-level feedback to select the model supplying each adopted update. Continuous image refinement and discrete language commitment operate on a shared generation state, while model-specific KV buffers provide attention context for selected queries. HybridStitch achieves 1.63--1.68$\times$ faster image generation and 1.15--1.58$\times$ diffusion large language model (dLLM) throughput on four main benchmarks. Estimated dLLM decode query cost decreases by 50--66% with comparable generation quality.

Subjects:	Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
Cite as:	arXiv:2603.07815 [cs.CV]
	(or arXiv:2603.07815v3 [cs.CV] for this version)
	https://doi.org/10.48550/arXiv.2603.07815 arXiv-issued DOI via DataCite

Submission history

From: Desen Sun [view email]
[v1] Sun, 8 Mar 2026 21:34:50 UTC (2,851 KB)
[v2] Tue, 13 Oct 2026 17:09:25 UTC (263 KB)
[v3] Wed, 14 Oct 2026 04:33:47 UTC (263 KB)

来源：arXiv cs.AI · arxiv.org