SANA-Video 2.0 5B 720p: NVIDIA's Efficient Video Model
SANA-Video 2.0 5B 720p renders 736x1280, 193-frame clips at 24 FPS from text or a first frame, with a 4-step DMD preview for text-to-video.
SANA-Video 2.0 5B 720p
Text-to-VideoImage-to-VideoLinear DiT720pNVIDIANVIDIA's 5B-class video diffusion transformer. The 50-step base checkpoint is post-trained for both text-to-video and text-image-to-video at 736x1280 for about eight seconds, and a 4-step DMD distillation is released alongside it as a text-to-video research preview.
| Developer | NVIDIA (SANA team) |
| Release Date | 2026-08 (base), 2026-08-27 (4-step preview) |
| Architecture | SanaVideo2_5B linear diffusion transformer, 32 layers, 4.47B parameters |
| Text Encoder | Gemma 2 2B IT |
| VAE | LTX 2.3, 128 latent channels, stride (8, 32, 32) |
| Output | 736x1280 at 193 frames and 24 FPS (about 8 seconds) |
| Generation Modes | Text-to-video, text-image-to-video |
| License | Apache-2.0 |
Architecture
SANA-Video 2.0 5B is a SanaVideo2_5B linear diffusion transformer with 32 layers, a hidden size of 2,560 and 4,466,980,960 trainable parameters (4.47B). Attention is hybrid: 75% gated bidirectional linear-attention layers carry the bulk of the sequence, while 25% periodic dense softmax-attention anchors restore global mixing, and shared timestep-independent Attention Residual aggregation runs every 8 layers. Text conditioning comes from google/gemma-2-2b-it, and the latent space follows the LTX 2.3 VAE contract with 128 latent channels and (8, 32, 32) temporal and spatial compression.
The released .pth is an inference artifact. The SANA-Video 2.0 EMA base weights and a ReFL post-training adapter were merged before release, so the file contains only the merged state_dict and no optimizer, scheduler or standalone LoRA tensors. The official entry point casts the transformer to BF16.
Output contract and sampling
The default 720p bucket is 736 x 1280 at 193 frames and 24 FPS, roughly eight seconds per clip. Both spatial dimensions must be divisible by 32, which is why the width is not the more familiar 1280 on its own, and frame counts must satisfy (num_frames - 1) % 8 == 0. The verified release example runs BF16 at CFG 8 with flow shift 12, 50 steps and motion score 20.
The 4-step preview is a full-model DMD distillation rather than a LoRA, and it is text-to-video only. It runs at CFG 1 with the sana_shift6_dpm sigma profile and does not use a truncated DPM-Solver trajectory: at each of the four fixed stages the model predicts velocity, computes x0, and re-noises at the next sigma (0.9998, 0.9472, 0.8569 and 0.6664) with a fresh draw from the same seeded generator. The preview emits 81 frames at 16 FPS, and 193 frames at 24 FPS in the model zoo listing.
Availability
Inference runs through the main branch of the Sana repository, which loads the checkpoint straight from the Hugging Face repository, and a Diffusers-format LTX 2.3 VAE has to be placed at output/pretrained_models/LTX-2.3-Diffusers/ or pointed to through vae.vae_pretrained. The model card and the model zoo list this Python path as the only supported route, and no ComfyUI integration for SANA-Video 2.0 is documented upstream.
Resources
Guides and workflows related to this model series.
Comments
Sign in with GitHub to join the discussion.