SANA-Video 2.0 5B 720p — 4-Step Preview

Research preview: This distilled checkpoint is an early T2V-only preview. For the original 50-step T2V + TI2V release, use SANA-Video_2.0_5B_720p.

Project page · Online demo · Documentation · GitHub · Paper

SANA-Video 2.0 is an efficient diffusion transformer for high-resolution video generation. This preview provides the latest global-step-1000 full-model EMA checkpoint distilled for four-stage text-to-video generation at 720p, plus an auxiliary RL LoRA for post-hoc strength adjustment. It combines gated bidirectional linear-attention layers with periodic dense softmax-attention anchors and shared Attention Residual aggregation.

Model details

Property	Value
Architecture	`SanaVideo2_5B`
Parameters	4,466,980,960 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
Task	Text-to-video only
Output buckets	736 × 1280; 81 frames / 16 FPS (5 seconds) or 193 frames / 24 FPS (8 seconds)
Model construction	Source tower at 480; explicit dynamic 736 × 1280 bucket
Text encoder	`google/gemma-2-2b-it`
VAE	LTX 2.3, 128 latent channels, stride `(8, 32, 32)`
Inference	BF16, CFG 1, 4 stages, `sana_shift6_dpm`
Online preview controls	Duration 5 or 8 seconds; Motion score 0–100; RL LoRA scale 0.3–1.0
License	Apache 2.0

Checkpoint lineage and format

The selected checkpoint is the latest global-step-1000 tri-temporal DMD EMA export initialized from the SANA-Video 2.0 SFT model after merging the ReFL step-500 adapter. DMD then updates the full transformer; the primary release artifact is therefore a full model, not a LoRA adapter.

The checkpoint contains a state_dict_ema tensor mapping only. It does not contain optimizer, learning-rate scheduler, gradient-scaler, or training-loop state. The inference entry point unwraps state_dict_ema, removes an optional model. key prefix, and casts the transformer to BF16.

The separately released rank-128 RL LoRA contains only inference metadata and the 256 pairs of lora_A/lora_B tensors. The online preview uses it to adjust the RL contribution around the full checkpoint's scale-1 reference:

W(s) = W(scale=1) + (s - 1) × (alpha / rank) × (B_RL @ A_RL)

Scale 1.0 exactly restores the uploaded full checkpoint. The Space defaults to 0.7 and exposes 0.3–1.0 in 0.1 increments. These lower scales are post-hoc adjustments around the same DMD checkpoint, not separately distilled models. Each change is reconstructed from immutable BF16 scale-1 references to avoid accumulating in-place rounding error.

Files

checkpoints/SANA_Video_2.0_5B_720p_4step.pth: distilled EMA transformer
adapters/SANA_Video_2.0_5B_720p_RL500_LoRA_rank128.pt: auxiliary RL scale controller
config.yaml: clean 5B source-tower config with the preview frame/FPS defaults
demo/: verified seed-4 MP4 and poster generated with the online defaults
LICENSE: Apache License 2.0

Verified 4-step example

This 1280 × 736, 81-frame, 16 FPS sample was generated from the released checkpoint with seed 4, Motion Score 20, and RL LoRA scale 0.7.

▶ Watch or download the verified four-step 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.

Four-stage sampling contract

This model does not use a truncated DPM-Solver trajectory. At each fixed stage, the model predicts velocity v, computes x0 = x - sigma * v, and—except at the terminal stage—re-noises x0 at the next sigma with a fresh draw from the same seeded generator.

Stage	Physical sigma	API timestep	Model embedding timestep
1	0.9998332262	999.8332262	999
2	0.9471688271	947.1688271	947
3	0.8568977118	856.8977118	856
4	0.6663702130	666.3702130	666

The rollout uses one BF16 initial-noise draw and three BF16 transition-noise draws. flow_shift is not applied by this sampler.

Online preview

The official Space uses the latest full checkpoint and auxiliary RL LoRA from this repository. Its public API takes the same inputs shown in the UI:

from gradio_client import Client

client = Client("Efficient-Large-Model/sana-video2-5b-720p-demo")
video_path, seed = client.predict(
    prompt="In a cozy, vintage room adorned with floral wallpaper, a cartoon rooster ...",
    video_duration="5 seconds",
    rl_lora_scale=0.7,
    motion_score=20,
    seed=4,
    randomize_seed=False,
    api_name="/generate",
)

Local inference

Use the current SANA main branch:

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 the config.

The following command runs the uploaded full checkpoint at its native RL scale 1.0. The Space applies the optional auxiliary adjustment when another scale is selected.

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_4step/checkpoints/SANA_Video_2.0_5B_720p_4step.pth \
  --txt_file=asset/samples/sana_video2_5b_720p_demo.txt \
  --task=t2v \
  --model.image_size=480 \
  --custom_height_width='[736,1280]' \
  --sampling_algo=fastvideo_dmd_4step \
  --generator_sigma_profile=sana_shift6_dpm \
  --cfg_scale=1.0 \
  --flow_shift=1.0 \
  --motion_score=20 \
  --negative_prompt=None \
  --num_frames=81 \
  --step=4 \
  --fps=16 \
  --seed=4 \
  --work_dir output/sana_video2_t2v_720p_4step_preview

The selected global-step-1000 checkpoint was evaluated on the full VBench T2V suite: 4,730 prompts across all 16 dimensions, using the same four-stage sampling contract, 736×1280 resolution, 81 frames, 16 FPS, and CFG 1.

Intended use

This model is intended for research, evaluation, and creative generation of short videos from text. It can also serve as a full-model starting point for further research under the license terms. It is not intended for first-frame image conditioning; use the base release for TI2V.

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

This is a research preview and has less extensive release validation than the original 50-step model.
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.
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.