Mainstream image and video generation models use VAE for multiple downsampling operations during the encoding stage, typically halving the resolution each time, so the original dimensions must be divisible by 8, 16, or 32 to avoid tensor dimension mismatches.