The present article addresses the modeling issues, related to the numerical simulation of aluminum depth filtration inside a real filter section in presence of non-negligible temperature variation. The filter structure is obtained by digitizing the computed tomography (CT) scanned data of a characteristic ceramic foam filter. The modeling takes the process conditions of a filtration trial in a pilot casting line into account. The incompressible flow of liquid aluminum is solved by the lattice Boltzmann method (LBM) and the heat transfer is solved by the finite volume method (FVM). The temperature dependences of the viscosity and the density of liquid metal are specifically considered. Although the consideration of buoyancy force significantly affects the fluid flow, the temperature-dependent viscosity plays only a minor role. A Lagrangian tracking of particles is performed for modeling the filtration of inclusions. In accordance with the depth filtration theory, the computational results confirm an exponential decrease in the particle concentration with the filter depth. The overall filtration efficiency is found to remain almost unaffected by both buoyancy and temperature-dependent viscosity. Using the numerically determined filtration coefficient, the filtration efficiency is extrapolated for a real filter of larger length, although the experimental data are underpredicted.