Definition

Temporal Coherence

Temporal coherence is making consecutive video frames belong together: no flicker, no morphing objects, no faces that subtly reshape. Video generation compresses footage into a latent space with a VAE, denoises it with a Diffusion Transformer whose attention spans both space and time, then decodes back to pixels. Holding identity, lighting, and physics across hundreds of frames is why video costs 100 to 1,000 times more compute than images.