Dietary habits are particularly important in fitness training, and a scientific fitness diet needs to be formulated according to the fitness user’s situation, fitness goals and exercise volume. In this paper, a multimodal personalized sports nutrition recommendation model incorporating users’ visual preferences is designed to address the important impact of visual features on the task of sports nutrition recommendation. The user’s visual preference is modeled using the Query-Key-Value attention mechanism, which extracts valuable visual information from their historical data and adds it to textual features. In the sports nutrition program generation part, guided by sports nutrition theory and based on the MMFV model, the sports nutrition program generation method was designed. Then, the multi-objective optimization problem of the sports nutrition scheme is planned and modeled, and a calorie-checking mechanism is added to the multi-objective particle swarm algorithm for the problem of calorie intake not meeting the fitness goal. According to the sports nutrition program recommendation algorithm to improve the three meals diets of the gym participants, through the analysis of relevant data to verify the scientificity of the recommended program. The average intake of protein, vitamin A, vitamin B1 and vitamin C of the fitness participants after the modification increased by 7.4%, 17.77%, 22.33% and 8.46%, respectively, compared with that before the modification. Scapular sebaceous thickness and abdominal sebaceous thickness were reduced by 2.2 mm and 1.47 mm, respectively, compared with the pre-modification period.TC, TG and LDL metabolism were in the normal range, which was significantly different from the pre-modification period (P