CASCADE: Context-Aware Relaxation for Speculative Image Decoding
Selin Yildirim, Subhajit Dutta Chowdhury, Mohammad Mahdi Kamani, Vikram Appia, Deming Chen, {seliny2, dchen}@illinois.edu, {sduttach, vikram.appia}@amd.com, [email protected] · 2026
Read CASCADE: Context-Aware Relaxation for Speculative Image Decoding by Selin Yildirim, Subhajit Dutta Chowdhury, Mohammad Mahdi Kamani, Vikram Appia, Deming Chen, {seliny2, dchen}@illinois.edu, {sduttach, vikram.appia}@amd.com, [email protected] (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-2605-07230.txt.
Description
Selin Yildirim and four co-authors speed up image generation by relaxing speculative decoding in a principled way. Image models generate step by step and stay slow even on top hardware. Draft-and-check schemes from text lose their gains because image models are so uncertain. Draft tokens get rejected too often to help.
The authors formalize two patterns in the target model's inner states: tokens that are interchangeable in meaning, and states that converge during prediction. CASCADE uses those redundancies to accept more drafts safely, with no extra training, and also feeds the signals into drafter training. Across image models, it reached leading speedups while keeping quality.
Catalog record
| CREATOR | Selin Yildirim, Subhajit Dutta Chowdhury, Mohammad Mahdi Kamani, Vikram Appia, Deming Chen, {seliny2, dchen}@illinois.edu, {sduttach, vikram.appia}@amd.com, [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.07230