Why it matters

Video generation is frontier. Understanding shapes capability trend.

Advertisement

The architecture

Latent video: temporal 3D VAE.

DiT with spatial + temporal attention.

Long generation: chained clips.

Video diffusionVideo latent3D VAEDiT-3Dspace + time attnDecodevideo framesSora, Kling, Veo lead; open models catching up (CogVideoX, Mochi)
Video diffusion.
Advertisement

How it works end to end

3D VAE: compress video to latents.

DiT with 3D attention.

Sampling: 25-50 steps per clip.

Long: consistency across chained clips.