Particle diffusometry (PD) is a technique of measuring the diffusion coefficient of a fluid sample by seeding it with tracer particles and observing their motion under a microscope. In microfluidic set-ups, the observed particles are often defocused and their motion is affected by factors such as fluid flow, which leads to high errors for conventional and deep learning-based PD (DPD) algorithms. This work improves the performance of DPD models by updating their architecture, avoiding temporal averaging in the input, and exploring the impact of various choices during training. These models provide state-of-the-art performance for generalised datasets regardless of particle shapes, concentration, flow or image noise and are called DPD-v2. These models provide a mean absolute error of 0.09$\mu$m2s−1 for Gaussian particles and 0.07$\mu$m2s−1 for defocused particles, which is 2x–4x lower errors as compared with the two following best methods. The performance of DPD-v2 models increases with crop size and the use of multiple stacks of images. The outputs of the DPD-v2 models were compared against the outputs from conventional algorithms on Gaussianised experimental no flow datasets, which provided < 0.5$\mu$m2s−1 mean absolute difference. Hence, the DPD-v2 models can be used in real-world scenarios.