Summary Back-projecting high-frequency (HF) waves is a common procedure for imaging rupture processes of large earthquakes (i.e. Mw > 7.0). However, obtained back-projection (BP) results could suffer from large uncertainties since high-frequency seismic waveforms are strongly affected by factors like source depth, focal mechanisms, and the Earth's 3D velocity structures. So far, these uncertainties have not been thoroughly investigated. Here, we use synthetic tests to investigate the influencing factors for which scenarios with various source and/or velocity set-ups are designed, using either Tohoku-Oki (Japan), Kaikoura (New Zealand), Java/Wharton Basin (Indonesia) as test areas. For the scenarios, we generate either 1D or 3D teleseismic synthetic data, which are then back-projected using a representative BP method, MUltiple SIgnal Classification (MUSIC). We also analyze corresponding real cases to verify the synthetic test results. The Tohoku-Oki scenario shows that depth phases of a point source can be back-projected as artifacts at their bounce points on the earth's surface, with these artifacts located far away from the epicenter if earthquakes occur at large depths, which could significantly contaminate BP images of large intermediate-depth earthquakes. The Kaikoura scenario shows that for complicated earthquakes, composed of multiple sub-events with varying focal mechanisms, BP tends to image sub-events emanating large amplitude coherent waveforms, while missing sub-events whose P nodal directions point to the arrays, leading to discrepancies either between BP images from different arrays, or between BP images and other source models. Using the Java event, we investigate the impact of 3D source-side velocity structures. The 3D bathymetry together with a water layer can generate strong and long-lasting coda waves, which are mirrored as artifacts far from the true source location. Finally, we use a Wharton Basin outer-rise event to show that the wavefields generated by 3D near trench structures contain frequency-dependent coda waves, leading to frequency-dependent BP results. In summary, our analyses indicate that depth phases, focal mechanism variations, and 3D source-side structures can affect various aspects of BP results. Thus, we suggest that target-oriented synthetic tests, for example, synthetic tests for subduction earthquakes using more realistic 3D source-side velocity structures, should be conducted to understand the uncertainties and artifacts before we interpret detailed BP images to infer earthquake rupture kinematics and dynamics.