FreqAdapter: Text-Guided Multi-Scale Frequency Adaptation for Fine-Tuning
In the arXiv paper 2605.08181, a novel fine-tuning approach known as FreqAdapter (Multi-Scale Frequency Adapter) is introduced. This method overcomes the shortcomings of current techniques by functioning within the frequency domain instead of the signal space, thereby minimizing information redundancy. FreqAdapter combines textual data and executes multi-scale fine-tuning of signals, employing a multi-scale adaptation strategy to enhance receptive fields across various frequency ranges. Its efficacy is validated through experiments on multimodal models like CLIP.
Key facts
- arXiv paper 2605.08181 introduces FreqAdapter
- FreqAdapter operates in the frequency domain
- It integrates textual information for fine-tuning
- It uses a multi-scale adaptation strategy
- Existing methods operate in signal space with information redundancy
- Existing methods use fixed prompts or adaptation layers
- FreqAdapter optimizes receptive fields across frequency ranges
- Experiments were conducted on multimodal models including CLIP
Entities
Institutions
- arXiv