SANA-Video 2.0 5B 720p

SANA-Video 2.0 is an efficient diffusion transformer for high-resolution video generation. This release provides the 5B-class checkpoint jointly post-trained for text-to-video (T2V) and text-image-to-video (TI2V) generation at 720p for about eight seconds.

The model combines gated bidirectional linear-attention layers with periodic dense softmax-attention anchors and shared Attention Residual aggregation. It uses Gemma 2 2B IT for text conditioning and the LTX 2.3 VAE contract with 128 latent channels and (8, 32, 32) temporal/spatial compression.

Model details

Property	Value
Architecture	`SanaVideo2_5B`
Parameters	4,466,980,960 trainable model parameters (4.47B)
Transformer	32 layers, hidden size 2,560
Attention	75% gated linear attention, 25% dense softmax anchors
Attention Residuals	Shared, timestep-independent aggregation every 8 layers
Tasks	Text-to-video and text-image-to-video
Output bucket	736 × 1280, 193 frames, 24 FPS (about 8 seconds)
Text encoder	`google/gemma-2-2b-it`
VAE	LTX 2.3, 128 latent channels, stride `(8, 32, 32)`
Recommended inference	BF16, CFG 8, flow shift 12, 50 steps, motion score 20
License	Apache 2.0

The checkpoint is an inference artifact containing only the merged model state_dict. It does not contain optimizer, scheduler, training-state, or standalone LoRA tensors. The EMA base weights and ReFL post-training adapter were merged before release. Stored tensors retain their merged source dtypes; the official inference entry point casts the transformer to BF16.

Files

checkpoints/SANA_Video_2.0_5B_720p.pth: merged transformer checkpoint
config.yaml: matching SANA training and inference configuration
LICENSE: Apache License 2.0

Verified release example

This sample was generated from the public checkpoint with seed 4. The encoded result is 1280 × 736, 193 frames, 24 FPS, and 8.04 seconds long.

▶ Watch or download the generated video

Prompt: In a cozy, vintage room adorned with floral wallpaper, a cartoon rooster sits comfortably in a floral-patterned armchair, sipping from a bottle of beer. The rooster, with its vibrant red comb and wattle, displays a range of expressions—smiling, nodding, and opening its beak wide in a cheerful manner. The setting includes wooden furniture and another beer bottle on the table, adding to the relaxed atmosphere. The camera captures the rooster from a close-up angle, emphasizing its animated movements and lively demeanor.

Inference

Support for this checkpoint is included in the main branch of NVlabs/Sana:

git clone https://github.com/NVlabs/Sana.git
cd Sana
bash environment_setup.sh sana
conda activate sana

Place the Diffusers-format LTX 2.3 VAE at output/pretrained_models/LTX-2.3-Diffusers/, or update vae.vae_pretrained in config.yaml.

Text-to-video

The command below is the exact command used for the verified release example:

bash inference_video_scripts/inference_sana_video.sh \
  --np 1 \
  --config configs/sana_video2/SanaVideo2_5B_720p.yaml \
  --model_path hf://Efficient-Large-Model/SANA-Video_2.0_5B_720p/checkpoints/SANA_Video_2.0_5B_720p.pth \
  --txt_file=asset/samples/sana_video2_5b_720p_demo.txt \
  --cfg_scale 8 \
  --flow_shift 12 \
  --step 50 \
  --fps 24 \
  --motion_score 20 \
  --seed 4 \
  --work_dir output/sana_video2_t2v_720p_demo

Text-image-to-video

Each line in asset/samples/sample_i2v.txt contains a prompt and an input-image path separated by <image>.

bash inference_video_scripts/inference_sana_video.sh \
  --np 1 \
  --config configs/sana_video2/SanaVideo2_5B_720p.yaml \
  --model_path hf://Efficient-Large-Model/SANA-Video_2.0_5B_720p/checkpoints/SANA_Video_2.0_5B_720p.pth \
  --txt_file=asset/samples/sample_i2v.txt \
  --task=ltx \
  --cfg_scale 8 \
  --flow_shift 12 \
  --step 50 \
  --fps 24 \
  --motion_score 20 \
  --work_dir output/sana_video2_ti2v_720p

The default 720p bucket is 736 × 1280 because both spatial dimensions must be divisible by 32. Frame counts must satisfy (num_frames - 1) % 8 == 0.

Intended use

This model is intended for research, evaluation, and creative generation of short videos from text, with optional first-frame image conditioning. It can also serve as a starting point for domain-specific fine-tuning under the license terms.

The model is not intended to produce factual evidence, identify people, make high-impact automated decisions, or generate content that violates privacy, copyright, applicable law, or platform policies.

Limitations and bias

Generated motion, anatomy, text rendering, object permanence, and physical interactions may be inconsistent, especially for crowded or highly dynamic scenes.
Prompt following can degrade for long, ambiguous, or compositionally complex instructions.
Image-conditioned generation can drift from fine details in the source image.
Outputs can reflect social and cultural biases present in training data and in the separately loaded text encoder.
The model does not independently verify whether generated content is factual, safe, or free of third-party rights.

Users should review outputs before publication, disclose synthetic media where appropriate, and add safeguards suited to their application.

Resources

Citation

If you use SANA-Video, please cite the SANA-Video work linked from the project page. SANA-Video 2.0-specific citation information will be added when it becomes available.