video diffusion: patchify, DiT, decode
raw video
T × H × W × 3 (240 frames @ 1080p)
3-D VAE encoder
spatiotemporal latent
patchify
t_p × h_p × w_p blocks
spatiotemporal DiT
factorized: spatial then temporal attn
cross-attn to T5-XXL text
same DDPM loss over spatiotemporal latents
L = E || ε - ε_θ( z_t, t, text, first_frame_opt ) ||²
flicker: independent per-frame sampling
per-frame noise is independent => jagged motion
coherent: joint sequence diffusion
shared noise + temporal attention => smooth motion