Sana Video 2

SANA-Video 2.0 is a video diffusion transformer from NVIDIA Research (Efficient AI Team and Singapore Lab), released as a project page and paper. It comes in 5B and 14B parameter sizes under one architecture and targets video generation up to 720p on a single GPU. It is the successor to earlier SANA models. Some sources refer to it as "Sana Video 2"; the official name is SANA-Video 2.0.

The architecture combines gated linear attention with periodic gated-softmax "anchor" layers at a 3:1 ratio. Block Attention Residuals pass block summaries into later linear layers. The models are trained from scratch rather than linearized from pretrained models. The authors report a VBench score of 84.30 at 480p in 13.2 s on one H100 (40 steps). They also report a compiled DiT forward pass 3.2× faster than a full-softmax baseline at 720p/60s. With Sol-Engine optimizations (kernel fusion, caching, sparse attention), the 5B pipeline reportedly reaches 13.06 s for a 5 s 720p clip, which the authors say is 120× faster than Wan 2.2-A14B on one H100.

The provided pages do not give a code, weights or demo link. A YouTube reviewer said it was not yet released, though earlier SANA models were, and noted visible artifacts in the samples. Availability, license and pricing are therefore unconfirmed.