Advances in Stylised Image Captioning: A Systematic Review of Methods, Benchmarks, Evaluation and Research Gaps

ABSTRACT

Image captioning has traditionally focused on factual, neutral descriptions, whereas stylised image captioning (SIC) generates captions that preserve visual content while expressing a specified style, such as humour, romance or sentiment. Despite growing interest, SIC's methods, datasets and evaluation practices remain fragmented. This systematic review synthesises methodological and evaluation trends across the literature. From 832 records (2016–2026), 49 primary studies were included after screening. The field progresses from CNN–LSTM and adversarial approaches towards Transformer‐based, controllable and large language model‐based generation, with substantial variation in how style is conditioned and supervised. About half the studies rely on paired stylised captions and two small English benchmarks from 2016 to 2017 remain central to evaluation. Fewer than half report human evaluation, with protocols differing in raters, scales and criteria. Evaluation is shifting from reference‐based n‐gram metrics towards classifier‐based style assessment, diversity, reference‐free similarity, preference‐based evaluation and LLM‐based judging, the latter appearing only from 2023 onward. The review identifies five persistent challenges: balancing stylistic expressiveness with visual grounding, standardising evaluation, extending beyond English and into cultural context, controlling safety, bias and hallucination and clarifying the boundary between stylised and controllable captioning. These findings motivate more diverse benchmarks and unified evaluation protocols.