CLIP bridges the gap between visual and language by learning both image and text representations simultaneously. As a large pre-trained visual language model, CLIP is highly generalisable and has demonstrated excellent few-shot learning capabilities. Numerous studies have been conducted on CLIP models for few-shot learning of downstream visual tasks, all of which have demonstrated excellent results. However, current research methods, such as the Adapter and Prompts methods, still fall short in extracting visual features for CLIP. Many methods only fine-tune the adapter after feature extraction, failing to fully utilise the feature extraction potential of CLIP. In addition, the fine-tuning approach using key-value cache improves performance significantly, but it requires careful tuning of the model’s hyperparameters for a specific dataset. Considering these issues, we propose a new approach: fine-tuning the multi-layer features with side adapters for adaptive selection in the visual backbone network. This approach efficiently extracts effective visual features for different layers of the task. Additionally, we propose augmenting the original features using dynamic feature fusion to reduce reliance on hyper-parameter tuning. Extensive experiments are conducted on 11 datasets to verify the superiority of the proposed method over existing methods.