Title: Polaris: Scaling Up Instruction-Guided Image Generation Towards Millions of Personalized Style Needs

Abstract: As users demand that image generation models rapidly adapt to highly varied and individualized preferences—such as creating visuals with unique stylistic traits or specific characteristics—traditional methods face significant hurdles. Conventional fine-tuning is often prohibitively expensive and challenging to scale. In response to these constraints, the research community has built an extensive repository of fine-tuned modules and adapters, with each component designed to address particular generation tasks and collectively forming a robust base for meeting emerging needs. This development prompts a critical question: rather than continuously training new models from scratch, can we systematically leverage this growing ecosystem to better satisfy user instructions? To answer this, we introduce Polaris, an intelligent retrieval framework that automatically identifies and combines appropriate models from the library according to user input. The core realization is that managing such a vast and diverse collection requires not only locating the most pertinent modules from thousands of options but also ensuring they are effectively aligned for instruction-based generation and editing. Polaris tackles this issue by indexing more than 6,500 checkpoints and 75,000 adapters, retrieving the most suitable components based on the user’s prompt. Consequently, it provides scalable, controllable, and well-aligned generation capabilities without the need for additional training.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC