AlignedNorm: Prompting Vision–Language Models via Coupled Prompt Field

Prompt learning has become an efficient way to adapt vision-language models (VLMs) to downstream tasks. However, existing end-to-end and decoupled methods often optimize base and new classes in isolated, task-specific feature spaces, which can lead to local optima and limited generalization.

We introduce AlignedNorm, a simple prompt-learning method built upon the concept of a Coupled Prompt Field. Instead of treating base and new classes independently, the coupled field places them in a shared optimization space where their learning dynamics mutually constrain each other. AlignedNorm realizes this coupling by dynamically aligning learnable prompts with the native feature scale of the pretrained VLM.

PDF | Project Page | 中译版 | Code | Official Version | Slides | BibTeX