Approve the Frame: Human Checkpoints and the Economics of Generative Video

Generative video is expensive to produce and still often wrong. A recent error audit found three or more visible errors in most generated clips, and the automatic judges that could filter them, from distribution metrics to multimodal language models, still fall short of human raters. At the same time, generating video from an explicit still image improves human-rated quality, and a still costs a small fraction of the energy and compute of a video. This paper argues that these findings point to one design: draw a cheap still first, let a person approve it, and only then pay for motion. We review evidence on generation errors, automatic evaluation, image-conditioned video, best-of-n selection and the energy cost of video, derive six design principles, and give a simple cost model showing that the saving grows with the share of candidates rejected at the still stage. We illustrate the principles with GetReelCom, whose workflow charges one credit for a still and ten for animating it, and does not charge failed generations. The paper reports no measurement of savings; it proposes an evaluation. Competing interests: the author is a co-founder of GetReelCom, the application used as the illustrative case. About the author: Amber Kesselaer is a growth consultant and an expert in marketing and AI video generation.