The article actually notes it takes about 15 minutes for a 5-second clip on an H100. Video diffusion steps are much heavier than single-image inference, so the

The article actually notes it takes about 15 minutes for a 5-second clip on an H100. Video diffusion steps are much heavier than single-image inference, so the low parameter count doesn't necessarily translate to low latency.