HybridStitch: Pixel and Timestep Level Model Stitching for Diffusion Acceleration
Desen Sun, Jason Hon 0, Jintao Zhang, Sihang Liu · 2026
Read HybridStitch: Pixel and Timestep Level Model Stitching for Diffusion Acceleration by Desen Sun, Jason Hon 0, Jintao Zhang, Sihang Liu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-07815.txt.
Description
Sun and colleagues present HybridStitch, a training-free framework using token-level feedback to select each adopted update's model. Stepwise diffusion generation needs varying model power at varying token spots, unlike prior mixes using cleaning step or drafting role. Ongoing image polish and discrete language choice share a state, while KV buffers provide context for picked queries. HybridStitch hits over 1.6 times faster image making and higher diffusion language model throughput. Decode query cost falls by about half to two-thirds at similar quality.
Catalog record
| CREATOR | Desen Sun, Jason Hon 0, Jintao Zhang, Sihang Liu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 32877 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.07815