Fluids and other deformable materials have highly mutable shapes, which are visibly influenced by both intrinsic properties (e.g. viscosity) and extrinsic forces (e.g. gravity, object interactions). How do we identify a liquid’s intrinsic properties across profound variations in shape caused by extrinsic factors? Previous findings suggest we are surprisingly good at matching viscosity across large variations in shape (“liquid constancy”). Here we ask which visual cues enable us to do this. Somehow the visual system abstracts features that are common to different instances of a liquid, while suppressing large differences in shape caused by extrinsic factors. In this study we tried to specify which geometric features observers use to achieve liquid constancy. We simulated eight variations of pouring liquids with seven different viscosities (‘test stimuli’). Each variation was influenced by a different noise force field, like gusts of wind that changed the way the liquid flowed, leading to substantial shape differences over time. Observers adjusted the viscosity of another variation (‘match stimulus’) until it appeared to be the same material as each test. We tested several time offsets to create volume differences between test and match stimuli. The experiment was performed with static and one-second moving stimuli. We find that observers show a high degree of constancy in matching the viscosity across the different variations. However, volume differences between test and match stimulus, especially with static stimuli, caused large effects of over- and under-estimation of viscosity. We then analyzed the 3D shapes of the samples to extract a wide range of shape measurements related to viscosity. We find that a number of cues related to curvatures, periodic movements of the liquids, and the way they spread out predict aspects of the observer’s performance, but that humans achieve better constancy than the cues predict.