Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion
Gang Dai, Yining Huang, Yiming Xia, Guohao Chen, Shuaicheng Niu · 2026
Read Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion by Gang Dai, Yining Huang, Yiming Xia, Guohao Chen, Shuaicheng Niu (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-2605-21907.txt.
Description
Gang Dai and co-authors propose RTS, a reward-guided method for scaling diffusion models at test time by steering the full denoising trajectory. Earlier methods only search preset noise candidates or explore the trajectory rigidly, so RTS finds key points where the denoising direction shifts. Tests on SD v3, FLUX, and Qwen-Image show GenEval gains of 20.7 percent, 15.6 percent, and 12.2 percent over baselines. Those key points cluster mainly in the middle stage, unlike the structure-sensitive early phase and the late attribute phase.
Catalog record
| CREATOR | Gang Dai, Yining Huang, Yiming Xia, Guohao Chen, Shuaicheng Niu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46013 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.21907