2.Self-selected HR region
We generate 720p videos on Wan2.1 14B from the text prompt only. Even early in denoising (k = 25), a linear head on the 13th DiT layer selects the HR region M without any region input; as shown below, the selected region (green) matches the visually important subject.
4.Dual-resolution attention
We fix all other factors and compare dual-resolution attention (ours) against a shared K/V cache on Wan2.1 1.3B (480p), trained for 5,000 steps. Due to the curse of mixed resolution, the shared K/V cache produces much stronger distortion along the seam.
5.Mixed-resolution decoder
We encode real videos at 480p (HR) and 240p (LR), keep the 480p latent inside a moving circle and the 240p latent outside, and compare our mixed-resolution decoder against resize-and-paste (which fills in each latent’s missing region, decodes them separately, and pastes the results together). Merging the two latents into a unified latent before decoding is critical to avoid distortion along the seam.
Limitation examples: fast motion
For fast-moving backgrounds, the transition between the low- and high-resolution regions can occasionally be noticeable.