Real-time and accurate estimation of live fish weight can play a critical role in precision feeding and growth monitoring, particularly in land land-intensive systems. However, environmental constraints, such as high stocking density, limited camera viewpoints, water turbidity, and fish movement, can often cause partial occlusion, where only the fish head is visible. Conventional visual measurement cannot fully meet the requirements of the large-scale production in modern aquaculture. In this study, two approaches were investigated to estimate the weight of largemouth bass (Micropterus salmoides) under head-visible and body-occluded conditions. The first approach was used to directly predict the fish weight using morphological features extracted from the visible head region, thus leveraging linear and nonlinear regression models trained on head-length and width measurements. While a straightforward and computationally efficient estimation, the absence of complete body information was introduced theled to significant errors under occlusion scenarios. The second approach was introduced with a two-stage generative completion framework. In the first stage, the partially visible head image was input into a generative model to synthesize the complete fish body, thus reconstructing previously unobserved morphological structures. The second stage was utilized the generated full-body image to extract the morphological features, including body length, height, and lateral area. Fish weight was then estimated using multivariate regression. Experimental evaluations were conducted on two datasets of largemouth bass. The two-stage approach was substantially improved the estimation accuracy. Specifically, the mean absolute error (MAE) decreased from an average of 83 to 63 g after the direct head estimation. Mean absolute percentage error (MAPE) was reduced from approximately 30% to 22%. The results indicate that the generative completion was effectively compensated for the missing morphological information caused by occlusion, thus providing quantitative weight estimates as a visual, interpretable reference for the reconstructed fish body. Compared with the direct estimation, this framework can offer a robust and intuitive solution to the biomass monitoring in intensive aquaculture, thus enhancing the reliability of growth assessment for the informed feeding decisions in the integration of computer vision into smart aquaculture. Furthermore, the generative stage can facilitate the visualization under extreme occlusion, allowing monitoring systems to assess fish morphology and size even when direct observation is partially obstructed. The findings underscore the potential of generative image completion with stereo vision regression in aquaculture environments. A promising direction can also offer for the real-time, non-invasive fish biomass estimation in an intelligent farm.