SANA-Video 2.0 4-Step Preview: NVIDIA's Fast 720p Video Model

Overview

The preview is a full-model DMD distillation, not a LoRA adapter: the global-step-1000 DMD EMA export was initialized from the SANA-Video 2.0 SFT model after merging a ReFL step-500 adapter, then DMD updates the entire transformer. The architecture combines 75% gated bidirectional linear-attention layers with 25% periodic dense softmax-attention anchors and shared Attention Residual aggregation every 8 layers. The text encoder is google/gemma-2-2b-it and the VAE is LTX 2.3 with 128 latent channels.

The verified seed-4 sample: a cartoon rooster in a floral vintage room, generated with the exact released checkpoint and settings. Source: SANA-Video 2.0 5B 720p 4-step

Four-Stage Sampling Contract

The model does not use a truncated DPM-Solver trajectory. At each fixed stage the model predicts velocity, computes x0, and re-noises at the next sigma with a fresh draw from the same seeded generator:

Inference runs BF16 at CFG 1 with the sana_shift6_dpm sigma profile and flow_shift disabled. The released checkpoint was evaluated on the full VBench T2V suite (4,730 prompts, all 16 dimensions) under these exact settings.

Availability

The checkpoint ships as a .pth EMA transformer plus a clean 5B source-tower config. Until the upstream PR merges, inference requires the preview branch of the Sana repository (feat/sana-video2-4step-preview) with a Diffusers-format LTX 2.3 VAE. The 50-step base release remains the entry point for image-conditioned TI2V: the distilled preview is text-to-video only and does not support first-frame image conditioning.

The verified four-step sample video (1280×736, 81 frames, 16 FPS), generated with seed 4. Source: SANA-Video 2.0 5B 720p 4-step