Movies taken by witnesses of extreme flood events are increasingly available on video sharing websites. They potentially provide highly valuable information on flow velocities and hydraulic processes that can help improve the post-flood determination of discharges in streams and flooded areas. We investigated the troubles and potential of applying the now mature large-scale particle image velocimetry (LSPIV) technique to such flood movies that are recorded under non-ideal conditions. Processing was performed using user-friendly, free software only, such as Fudaa-LSPIV. Typical issues related to the image processing and to the hydrological analysis are illustrated using a selected example of a pulsed flash-flood flow filmed in a mountainous torrent. Simple corrections for lens distortion (fisheye) and limited incoherent camera movement (shake) were successfully applied, and the related errors were reduced to a few percents. Testing the different image resolution levels offered by YouTube showed that the difference in time-averaged longitudinal velocity was less than 5% compared with full resolution. A limited number of GRPs, typically 10, is required, but they must be adequately distributed around the area of interest. The indirect determination of the water level is the main source of uncertainty in the results, usually much more than errors because of the longitudinal slope and waviness of the free-surface of the flow. The image-based method yielded direct discharge estimates of the base flow between pulses, of the pulse waves, and of the time-averaged flow over a movie sequence including a series of five pulses. A comparison with traditional indirect determination methods showed that the critical-depth method may produce significantly biassed results for such a fast, unsteady flow, while the slope-area method seems to be more robust but would overestimate the time-averaged flow rate if applied to the high-water marks of a pulsed flow.