Variable Resolution

Key Insight

Most diffusion models are welded to the one resolution they trained on; Sora's headline claim was variable resolution — one model that generates clips at many sizes, durations, and aspect ratios. It works because a DiT processes a sequence of spatiotemporal patches rather than a fixed grid, and 3D RoPE encodes each token's position as a rotation that extrapolates gracefully to lengths and shapes never seen in training. By feeding the model more or fewer tokens you get a taller, wider, or longer video from the very same weights. Training across aspect-ratio buckets and then testing on an aspect ratio it never saw is the concrete experiment that proves the claim — or exposes where it breaks down.