Text-to-Video

Text-to-video extends generative imaging across time, and the added dimension is genuinely difficult. Every frame must be plausible individually *and* consistent with its neighbors: objects keep their identity, lighting stays coherent, motion obeys something like physics.

Progress has been rapid, with clip length, resolution and temporal stability all improving, and synchronized audio generation now appearing alongside. Output remains short-form, and the reliable workflow is generating many candidates and selecting, rather than directing a specific shot.

Compute cost is the practical constraint — video generation is dramatically more expensive per second of output than image generation, which shapes both pricing and how much iteration is affordable.

Video is the most expensive media format to produce conventionally, so the cost collapse has the largest proportional effect here. It also raises the deepfake stakes: convincing synthetic video of real people is now achievable without specialist skill.

Budget for iteration — you will generate many clips per usable one. Treat it as a source of B-roll, concepts and short segments rather than as a way to direct a specific scene.