jaayeon/agsm

AGSM is the code repository for a research paper that the authors describe as a 'lightweight, reward-free post-training method' for improving how well text-to-image diffusion models follow the prompt they were given. The work was accepted to the 2026 International Conference on Machine Learning, the conference known as ICML, and the authors say it was selected for the 'Spotlight' track, which is a curated subset of accepted papers. The four authors are listed as Jaa-Yeon Lee, Yeobin Hong, Taesung Kwon, and Jong Chul Ye, all from KAIST in South Korea. In plain terms, text-to-image diffusion models are the family of AI systems that generate a picture from a sentence, such as Stable Diffusion and similar tools. A common complaint with these models is that the picture they produce often misses parts of the prompt, for example skipping one of the objects, getting a colour wrong, or putting the wrong number of items in the frame. The authors call this the 'text-image alignment' problem, and AGSM is their proposed fix. The method is described as 'lightweight' at 1.8 million parameters, which is small compared with the underlying diffusion model. It is also described as 'post-training', meaning it is applied on top of a model that has already been trained, instead of training a new model from scratch. And it is 'reward-free', meaning it does not need a separate scoring or reward model to grade the generated images, which is the route many recent alignment methods take. Beyond these three descriptors the README does not explain how the method works; the technical detail lives in the paper itself. The README is otherwise sparse. It links to the arXiv preprint and to a project page at jaayeon.github.io/AGSM, and it shows a header image with example outputs. The setup section gives the git clone command, but the install instructions are marked 'Coming Soon', so the running code is not yet usable from this repository at the time of writing. There is a short acknowledgements note saying the readme style was modelled on another open-source project called FlashWorld.