🎬 SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer¶

🎬 Demos of SANA-Video¶

📽️ About SANA-Video¶

SANA-Video is a small diffusion model designed for efficient video generation, capable of synthesizing high-resolution videos (up to $720 \times 1280$) and minute-length duration with strong text-video alignment, while maintaining a remarkably fast speed.It enables low-cost, high-quality video generation and can be deployed efficiently on consumer GPUs like the RTX 5090.

SANA-Video's Core Contributions:

Efficient Architecture (Linear DiT): Leverages linear attention as the core operation, which is significantly more efficient than vanilla attention for video generation due to the large number of tokens processed.
Long-Sequence Capability (Constant-Memory KV Cache): Introduces a Constant-Memory KV cache for Block Linear Attention. This block-wise autoregressive approach uses a fixed-memory state derived from the cumulative properties of linear attention, which eliminates the need for a traditional KV cache, enabling efficient minute-long video generation.
Low Training Cost: Achieved effective data filters and model training strategies, narrowing the training cost to only 12 days on 64 H100 GPUs, which is just 1% of the cost of MovieGen.
State-of-the-Art Speed and Performance: Achieves competitive performance compared to modern SOTA small diffusion models (e.g., Wan 2.1-1.3B) while being $16\times$ faster in measured latency。Deployment Acceleration: Can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s ($2.4\times$ speedup).

In summary, SANA-Video enables high-quality video synthesis at an unmatched speed and low operational cost.

💻 Block Causal Linear Attention && Causal Mix-FFN Mechanism¶

🏃 How to Inference¶

1. How to use Sana-Video Pipelines in `🧨diffusers`¶

Note

Upgrade your diffusers to use SanaVideoPipeline:

pip install git+https://github.com/huggingface/diffusers

Text-to-Video: SanaVideoPipeline¶

import torch
from diffusers import SanaVideoPipeline
from diffusers import AutoencoderKLWan
from diffusers.utils import export_to_video

model_id = "Efficient-Large-Model/SANA-Video_2B_480p_diffusers"
pipe = SanaVideoPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
pipe.vae.to(torch.float32)
pipe.text_encoder.to(torch.bfloat16)
pipe.to("cuda")
motion_score = 30

prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest, golden light glimmers on his hair as sunlight filters through the leaves. He wears a light shirt, wind gently blowing his hair and collar, light dances across his face with his movements. The background is blurred, with dappled light and soft tree shadows in the distance. The camera focuses on his lifted gaze, clear and emotional."
negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
motion_prompt = f" motion score: {motion_score}."
prompt = prompt + motion_prompt

video = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    height=480,
    width=832,
    frames=81,
    guidance_scale=6,
    num_inference_steps=50,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(video, "sana_video.mp4", fps=16)

Image-to-Video: SanaImageToVideoPipeline¶

import torch
from diffusers import SanaImageToVideoPipeline, FlowMatchEulerDiscreteScheduler
from diffusers.utils import export_to_video, load_image

pipe = SanaImageToVideoPipeline.from_pretrained("Efficient-Large-Model/SANA-Video_2B_480p_diffusers")
# pipe.scheduler = FlowMatchEulerDiscreteScheduler(shift=pipe.scheduler.config.flow_shift)
pipe.transformer.to(torch.bfloat16)
pipe.text_encoder.to(torch.bfloat16)
pipe.vae.to(torch.float32)
pipe.to("cuda")

motion_score = 30
prompt = "A woman stands against a stunning sunset backdrop, her , wavy brown hair gently blowing in the breeze. She wears a veless, light-colored blouse with a deep V-neckline, which ntuates her graceful posture. The warm hues of the setting sun cast a en glow across her face and hair, creating a serene and ethereal sphere. The background features a blurred landscape with soft, ing hills and scattered clouds, adding depth to the scene. The camera ins steady, capturing the tranquil moment from a medium close-up e."
negative_prompt = "A chaotic sequence with misshapen, deformed limbs eavy motion blur, sudden disappearance, jump cuts, jerky movements, d shot changes, frames out of sync, inconsistent character shapes, oral artifacts, jitter, and ghosting effects, creating a disorienting al experience."
motion_prompt = f" motion score: {motion_score}."
prompt = prompt + motion_prompt

image = load_image("https://raw.githubusercontent.com/NVlabs/Sana//heads/main/asset/samples/i2v-1.png")

output = pipe(
    image=image,
    prompt=prompt,
    negative_prompt=negative_prompt,
    height=480,
    width=832,
    frames=81,
    guidance_scale=6,
    num_inference_steps=50,
    generator=torch.Generator(device="cuda").manual_seed(42),
).frames[0]

export_to_video(output, "sana-ti2v-output.mp4", fps=16)

2. Inference with TXT file¶

Text-to-Video¶

bash inference_video_scripts/inference_sana_video.sh \
      --np 1 \
      --config configs/sana_video_config/Sana_2000M_480px_AdamW_fsdp.yaml \
      --model_path hf://Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth \
      --txt_file=asset/samples/video_prompts_samples.txt \
      --cfg_scale 6 \
      --motion_score 30 \
      --flow_shift 8 \
      --work_dir output/sana_t2v_video_results

Image-to-Video¶

bash inference_video_scripts/inference_sana_video.sh \
      --np 1 \
      --config configs/sana_video_config/Sana_2000M_480px_AdamW_fsdp.yaml \
      --model_path hf://Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth \
      --txt_file=asset/samples/sample_i2v.txt \
      --task=ltx \
      --cfg_scale 6 \
      --motion_score 30 \
      --flow_shift 8 \
      --work_dir output/sana_ti2v_video_results

💻 How to Train¶

# 5s Video Model Pre-Training
bash train_video_scripts/train_video_ivjoint.sh \
      configs/sana_video_config/Sana_2000M_480px_AdamW_fsdp.yaml \
      --data.data_dir="[data/toy_data]" \
      --train.train_batch_size=1 \
      --work_dir=output/sana_video \
      --train.num_workers=10 \
      --train.visualize=true

Convert pth to diffusers safetensor¶

python scripts/convert_scripts/convert_sana_video_to_diffusers.py --dump_path output/SANA_Video_2B_480p_diffusers --save_full_pipeline

Performance¶

VBench Results - 480p Resolution¶

Text-to-Video¶

Methods	Latency (s)	Speedup	#Params (B)	Total ↑	Quality ↑	Semantic / I2V ↑
MAGI-1	435	1.1×	4.5	79.18	82.04	67.74
Step-Video	246	2.0×	30	81.83	84.46	71.28
CogVideoX1.5	111	4.4×	5	82.17	82.78	79.76
SkyReels-V2	132	3.7×	1.3	82.67	84.70	74.53
Open-Sora-2.0	465	1.0×	14	84.34	85.4	80.72
Wan2.1-14B	484	1.0×	14	83.69	85.59	76.11
Wan2.1-1.3B	103	4.7×	1.3	83.31	85.23	75.65
SANA-Video	60	8.0×	2	84.17	84.85	81.46

Image-to-Video¶

Methods	Latency (s)	Speedup	#Params (B)	Total ↑	Quality ↑	Semantic / I2V ↑
MAGI-1	435	1.1×	4.5	89.28	82.44	96.12
Step-Video-TI2V	246	2.0×	30	88.36	81.22	95.50
CogVideoX-5b-I2V	111	4.4×	5	86.70	78.61	94.79
HunyuanVideo-I2V	210	2.3×	13	86.82	78.54	95.10
Wan2.1-14B	493	1.0×	14	86.86	80.82	92.90
SANA-Video	60	8.2×	2	88.02	79.65	96.40

VBench Results - 720p Resolution¶

Models	Latency (s)	Total ↑	Quality ↑	Semantic ↑
Wan-2.1-14B	1897	83.73	85.77	75.58
Wan-2.1-1.3B	400	83.38	85.67	74.22
Wan-2.2-5B	116	83.28	85.03	76.28
SANA-Video-2B	36	84.05	84.63	81.73

Summary: Compared with the current SOTA small video models, SANA's performance is very competitive and speed is much faster. SANA provides 83.71 VBench overall performance with only 2B model parameters, 16× acceleration at 480p, and achieves 84.05 total score with only 36s latency at 720p resolution.

VBench Results - 30s Long Video Vbench¶

Models	FPS	Total ↑	Quality ↑	Semantic ↑
SkyReels-V2	0.49	75.29	80.77	53.37
FramePack	0.92	81.95	83.61	75.32
Self-Forcing	17.0	81.59	83.82	72.70
LongSANA-2B	27.5	82.29	83.10	79.04

Summary: Compared with the current SOTA long video generation models, LongSANA (SANA-Video + LongLive)'s speed and performance is very competitive. LongSANA's 27FPS generatin speed on H100 makes real-time generation possible.

Citation¶

@misc{chen2025sana,
      title={SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer},
      author={Chen, Junsong and Zhao, Yuyang and Yu, Jincheng and Chu, Ruihang and Chen, Junyu and Yang, Shuai and Wang, Xianbang and Pan, Yicheng and Zhou, Daquan and Ling, Huan and others},
      year={2025},
      eprint={2509.24695},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.24695},
}

🎬 SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer¶

🎬 Demos of SANA-Video¶

📽️ About SANA-Video¶

💻 Block Causal Linear Attention && Causal Mix-FFN Mechanism¶

🏃 How to Inference¶

1. How to use Sana-Video Pipelines in 🧨diffusers¶

Text-to-Video: SanaVideoPipeline¶

Image-to-Video: SanaImageToVideoPipeline¶

2. Inference with TXT file¶

Text-to-Video¶

Image-to-Video¶

💻 How to Train¶

Convert pth to diffusers safetensor¶

Performance¶

VBench Results - 480p Resolution¶

Text-to-Video¶

Image-to-Video¶

VBench Results - 720p Resolution¶

VBench Results - 30s Long Video Vbench¶

Citation¶

1. How to use Sana-Video Pipelines in `🧨diffusers`¶