Headline generation for news articles involves crafting brief, impactful summaries of lengthy texts, an essential task for digital content as it grabs readers’ attention and quickly communicates the main message. Although significant progress has been made for widely spoken languages like English, considerable challenges persist in low-resource settings, particularly for many Indian languages, due to the lack of high-quality annotated data and the complexity of their linguistic structures. This research addresses the challenges of headline generation for low-resource languages, with a focus on Tamil, by comparing traditional fine-tuning methods to Low-Rank Adaptation (LoRA) techniques to enhance model efficiency using the IndicBART-XLSum and IndicBARTSS models. Our approach involved data preprocessing, sequence tagging, and developing custom evaluation metrics. We first fine-tuned the models using traditional methods, then optimized the process with LoRA to reduce the number of trainable parameters. The results demonstrated that traditional fine-tuning delivered the best performance, with IndicBART-XLSum achieving a BLEU score of 0.475 and a BERTScore F1 of 0.960, indicating strong accuracy and coherence in headline generation. LoRA significantly reduced computational costs and training time by over 50% but came with a slight performance trade-off, yielding a BLEU score of 0.426 and a BERTScore F1 of 0.954. This study highlights the trade-off between efficiency and accuracy in NLP applications for low-resource languages. While LoRA substantially reduces training time, traditional fine-tuning achieves better performance. Our findings suggest that optimizing NLP models for low-resource Indic languages requires a balance of computational efficiency and output quality.