2025-12-31 上海交通大学 (SJTU)
＜関連情報＞
- https://global.sjtu.edu.cn/en/news-events/news/2458
- https://www.science.org/doi/abs/10.1126/science.adv7434
大規模インテリジェントセマンティックビジョン生成のための全光合成チップ All-optical synthesis chip for large-scale intelligent semantic vision generation
Yitong Chen https://orcid.org/0000-0003-4207-4959, Xinyue Sun https://orcid.org/0009-0008-0529-4323, Longtao Tan, Yizhou Jiang, […] , and Guangtao Zhai
Science Published:18 Dec 2025
DOI:https://doi.org/10.1126/science.adv7434
Editor’s summary
Despite extensive research on optical computing in recent years, there has been limited progress in directly addressing leading-edge artificial intelligence (AI) practical applications. This is especially true for large-scale models, which have the potential to radically change society but at the same time face severe shortages of computing power and energy. By integrating millions of photonic neurons and varying the network dimension at light speed through proposed optical latent space and Bayes-based training algorithms, Chen et al. developed an all-optical chip that was able to experimentally implement large-scale cutting-edge AI tasks of generative models and semantic manipulation of complicated colorful images with impressive speed and energy efficiency compared with state-of-the-art electronic and photonic chips. —Yury Suleymanov
Abstract
Large-scale generative artificial intelligence (AI) is facing a severe computing power shortage. Although photonic computing achieves excellence in decision tasks, its application in generative tasks remains formidable because of limited integration scale, time-consuming dimension conversions, and ground-truth-dependent training algorithms. We produced an all-optical chip for large-scale intelligent vision generation, named LightGen. By integrating millions of photonic neurons on a chip, varying network dimension through proposed optical latent space, and Bayes-based training algorithms, LightGen experimentally implemented high-resolution semantic image generation, denoising, style transfer, three-dimensional generation, and manipulation. Its measured end-to-end computing speed and energy efficiency were each more than two orders of magnitude greater than those of state-of-the-art electronic chips, paving the way for acceleration of large visual generative models.
