Esta nota todavía no está traducida, así que se muestra la fuente en inglés.
Video generation
Video generation is not just many images. It must model motion, continuity, physics, camera behavior, identity persistence, editing intent, and audio-visual alignment over time.
What video adds
| Challenge | Why it is hard |
|---|---|
| Temporal consistency | objects and identities must persist across frames |
| Motion | generated dynamics need plausible trajectories |
| Camera control | viewpoint, zoom, cuts, and pans affect meaning |
| Long context | coherence degrades as duration grows |
| Cost | frames multiply compute, memory, and storage |
| Evaluation | quality depends on motion, story, artifacts, and safety |
Common workflows
- Text-to-video from a prompt.
- Image-to-video from a starting frame.
- Video-to-video style or subject transformation.
- Inpainting and editing within a clip.
- Storyboard or keyframe-conditioned generation.
- Synthetic training data for perception tasks.
Product constraints
Video systems often need stronger review than images: likeness rights, deepfake risk, brand safety, misinformation, watermarking, and provenance. They also need clear expectations about duration, resolution, editability, and render time.
Pitfall
Short cherry-picked clips can hide instability. Evaluate repeated generations, motion continuity, temporal artifacts, prompt adherence, and safety across a task suite.
Connects to: diffusion models · generative media eval · cost optimization