Video generators such as Sora use a Diffusion Transformer (DiT): video is split into a grid of spacetime patches, and a transformer denoises all of them together, step by step, from pure noise:
x_(t-1) = 1/√α_t · ( x_t − (1−α_t)/√(1−ᾱ_t) · ε_θ(x_t, t, c) ) + σ_t·z
ε_θ(x_t,t,c) = ε_uncond + w · (ε_cond − ε_uncond) // classifier-free guidance, w = CFG scale
coherence = 1 − mean_f | frame_f − frame_(f-1) | / range
- Denoising steps — how many reverse-diffusion steps the filmstrip takes to go from noise to a clean clip; more steps reveal detail more gradually, exactly like a real video-diffusion sampler.
- Guidance scale (CFG) — classifier-free guidance strength w in the formula above; higher values push every frame's colours further from the noise average toward a sharper, higher-contrast target.
- Temporal coherence — how strongly each frame's per-patch reveal-schedule is shared with its neighbours (a stand-in for DiT's spacetime attention); at 0% every frame denoises independently and the clip flickers, at 100% all frames reveal in lockstep.
- Motion speed — how fast the underlying scene content (the thing being generated) travels across the temporal window, independent of the denoising schedule.