The fastest-moving frontier

Text-to-video is the newest and most rapidly improving corner of generative AI — what looks state-of-the-art this quarter is routinely beaten the next. Tools like Runway, Pika, Luma Dream Machine, Kling, Hailuo, and OpenAI's Sora (plus Google's Veo) can turn a text prompt or a starting image into short, increasingly coherent clips with convincing motion.

Two ways in

• Text-to-video — describe a scene and get a clip. Best for conjuring something from nothing.
• Image-to-video — start from a still (often one you generated) and animate it. This gives you far more control, because you've already locked the look; the model only has to add motion.

What it's genuinely good for today

• Short B-roll and social clips where perfect continuity isn't required.
• Animating a still — bringing a photo or illustration to life.
• Avatar presenters (Synthesia, HeyGen, D-ID) that turn a script into a person-on-camera video, with translation and lip-sync — popular for training and marketing.

The current limits

Be realistic about what's hard. Consistency is the main constraint: keeping a character, object, or style identical across a long clip or several shots is still unreliable, and fine details (hands, text, faces over time) can warp. The practical workaround is to keep clips short (a few seconds), use concrete prompts, generate several takes, and edit the best moments together rather than expecting one perfect long shot.

Today's text-to-video shines in short bursts, not long takes. Think of it as a generator of clips you assemble, not a one-click filmmaker — and plan your prompts and editing around that.

Try this: Take a still image — generated or a photo — and run it through an image-to-video tool with a simple motion prompt ("slow zoom in, gentle breeze"). Starting from a fixed image shows you how much more control image-to-video gives than starting from text alone.