SANA-Video: NVIDIA's Linear Diffusion Transformer for Video

Variants

Select a variant above for curated weight downloads, tutorials, and related information.

About the SANA-Video line

SANA-Video extends NVIDIA's SANA linear diffusion transformer from images to video. Self-attention is replaced by linear attention, which removes the quadratic cost that dominates token-heavy video generation, and the design adds rotary position embeddings plus a 1D temporal convolution inside the Mix-FFN so that a pre-trained image model can be adapted to video by continued pre-training. A block linear attention module with a constant-memory KV cache keeps the context of every previous block at a fixed memory cost, which is what makes minute-length generation practical. The authors report a training budget of 12 days on 64 H100 GPUs.

SANA-Video 2.0 is the current generation. The 5B 720p checkpoint mixes 75% gated bidirectional linear-attention layers with 25% periodic dense softmax-attention anchors, and aggregates shared Attention Residuals every 8 layers. It has 32 layers, a hidden size of 2,560 and 4.47B trainable parameters. Conditioning uses Gemma 2 2B IT as the text encoder, and decoding uses the LTX 2.3 VAE contract with 128 latent channels and (8, 32, 32) temporal and spatial compression. The checkpoint is post-trained with ReFL for both text-to-video and text-image-to-video, so a first frame can be supplied as an extra condition.

The model hub exposes both released weights as one variant. The 50-step 720p base checkpoint is the entry point for joint text-to-video and text-image-to-video generation, and the 4-step DMD distillation is distributed as a separate research preview that is text-to-video only. The official model zoo also announces a 14B 480p model as coming soon, which is not released yet.

Availability

Upstream documents the main branch of the Sana repository as the inference entry point and expects a Diffusers-format LTX 2.3 VAE to be provided locally. The model card and the model zoo do not document a ComfyUI node pack for SANA-Video 2.0.