Text-to-image generation exploded in just two years. GANs like VQGAN-CLIP and GauGAN2 kicked things off, then DALL-E arrived in early 2021, DALL-E 2 followed in April 2022, and diffusion models — Stable Diffusion, Imagen — changed the game entirely. DreamStudio, RunwayML GEN-1, Midjourney. The whole ecosystem matured at a pace that felt almost reckless.

Text-to-video hasn’t had that moment. Not even close.

The core problem is structural. Video models have to maintain spatial and temporal consistency across frames, which creates long-term dependencies that make training computationally brutal — unaffordable for most researchers, as the article puts it. Then there’s the data problem. Multi-modal video datasets are scarce, sparsely annotated, and often too vague to teach a model anything about complex movement semantics. And the captioning question remains wide open: a single short prompt can’t describe what happens over time. You need a sequence of prompts, a narrative arc.

Early work leaned on GANs and VAEs — Text2Filter, TGANs-C — generating frames auto-regressively at low resolutions with isolated motions. Then came the transformer wave: Phenaki, Make-A-Video, NUWA, VideoGPT, CogVideo. Phenaki stands out because it can generate arbitrarily long videos conditioned on a story line, while NUWA-Infinity proposes an autoregressive over autoregressive mechanism for infinite generation. TATS went hybrid, pairing VQGAN for images with a time-sensitive transformer for sequential frames.

But here’s the thing the article hints at without saying outright: the gap between image and video generation isn’t just technical, it’s economic. Stable Diffusion succeeded partly because it was cheap enough for hobbyists to run on consumer GPUs. Text-to-video models still require sliding window approaches to generate anything longer than a few seconds, which makes deployment a headache. Until that compute cost drops dramatically, text-to-video will remain a research toy rather than a production tool — the same place text-to-image was in early 2021, right before DALL-E broke things open.