The "first" video generation model already exists. It all hangs on "reasonable quality", which I don't know what to make of.

An H100 costs ~$1.89/hour, which is ~3min for $0.1, so if a H100 can produce 1min of video (60s*30fps=1800 frames) in ~3min of GPU time (~10fps), then it'd resolve NO.

How these frames are rendered, at what resolution, whether upscaling or frame interpolation is used, etc will have a big impact. Makes "reasonable quality" a difficult measure, but I suppose 720p 30fps with frame interpolation every other frame (15fps raw) and 1.5x upscaling (from 480p) might count? Which capabilities it needs to have to count is still unclear.

Any model? Runway Gen-2?