InstructEngine: Instruction-driven Text-to-Image Alignment
Xingyu Lu 1,2,† Yuhang Hu 2,3,† YiFan Zhang 2,3 Kaiyu Jiang 2 Changyi Liu 2 Tianke Zhang 2 Jinpeng Wang 1 Chun Yuan 1,‡ Bin Wen 2,‡ ♠ Fan Yang 2 Tingting Gao 2 Di Zhang 2 1 Tsinghua University 2 Kuaishou Technology 3 CASIA † Equal Contribution ‡ Corresponding Author ♠ Project Leader · 2025
Read InstructEngine: Instruction-driven Text-to-Image Alignment by Xingyu Lu 1,2,† Yuhang Hu 2,3,† YiFan Zhang 2,3 Kaiyu Jiang 2 Changyi Liu 2 Tianke Zhang 2 Jinpeng Wang 1 Chun Yuan 1,‡ Bin Wen 2,‡ ♠ Fan Yang 2 Tingting Gao 2 Di Zhang 2 1 Tsinghua University 2 Kuaishou Technology 3 CASIA † Equal Contribution ‡ Corresponding Author ♠ Project Leader (2025) 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-2504-10329.txt.
Description
Xingyu Lu and a large team present InstructEngine in a 2025 paper, a way to align text-to-image models with what users ask for. Feedback-based training is the standard path to that match, but it rests on hand-labeled likes and dislikes. Those labels cost a lot, and a reward model in the loop adds compute with no sure accuracy.
Most past methods also waste the text: they read the image score alone as the signal, which is thin. InstructEngine is driven by the instruction itself. It aims to give a richer, cheaper signal for how well each image fits its prompt.
Catalog record
| CREATOR | Xingyu Lu 1,2,† Yuhang Hu 2,3,† YiFan Zhang 2,3 Kaiyu Jiang 2 Changyi Liu 2 Tianke Zhang 2 Jinpeng Wang 1 Chun Yuan 1,‡ Bin Wen 2,‡ ♠ Fan Yang 2 Tingting Gao 2 Di Zhang 2 1 Tsinghua University 2 Kuaishou Technology 3 CASIA † Equal Contribution ‡ Corresponding Author ♠ Project Leader |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-10329