1. Quick start
Follow the SGLang Diffusion installation guide for your GPU platform.2. Model capabilities
SANA-Video 2B 480p is a lightweight text-to-video model served through a native SGLang Diffusion pipeline. Its main advantage is deployment cost: at 2B parameters it serves 480p video from a single GPU with a plainsglang serve invocation, no
parallelism flags required.
The released generation profile produces 832×480 output at 81 frames and
16 FPS over 50 inference steps; a compact 17-frame, 8-step profile is covered
by server CI for quick validation. Motion strength can be steered directly
from the prompt with an optional motion score: N. suffix.
3. Serve SANA-Video
Command
4. Generate a video
The following request uses the compact 17-frame, 8-step profile covered by server CI. Use the model defaults of 81 frames and 50 steps for the released generation profile.Python
5. Request constraints
- The default profile uses
832x480, 81 frames, 50 inference steps, and 16 FPS. - Frame counts are aligned to
4n+1; for example, a request for 80 frames is adjusted to 77. - Use width and height values divisible by 16.
- The prompt supports an optional
motion score: N.suffix to express the desired amount of motion.
6. SANA-Video 2.0
The separateEfficient-Large-Model/SANA-Video_2.0_5B_720p checkpoint supports
text-to-video and first-frame image conditioning. Its default profile is
1280×736, 193 frames at 24 FPS, 50 steps, and guidance scale 8. This integration
loads the official checkpoint directly and uses Gemma 2 and the LTX 2.3 VAE.
On a single NVIDIA CUDA GPU, generate a short video with:
Command
--image-path first_frame.png to condition the video on an image. Frame
counts align to 8n+1; width and height must be divisible by 32. The 4-step
preview is a different checkpoint and is not supported by this pipeline.
For repeated requests at a fixed shape on an NVIDIA CUDA GPU with enough memory
to keep all components resident, capture breakable CUDA graphs at startup:
Command
--quality high to run the eight softmax anchor layers with BF16-input
flash attention instead of the FP32 memory-efficient kernel the default path
uses. Accumulation stays FP32 inside the kernel, but the output is no longer
bit-identical to the default path.
With two GPUs, add --num-gpus 2 to split the video tokens across them
(sequence parallelism). The result is not bit-identical to the single-GPU
result. The sequence-parallel degree must divide the 10 softmax attention
heads; --ring-degree above 1 and --enable-cfg-parallel are not supported.
For offline weights, keep the model directory name, such as
/path/to/SANA-Video_2.0_5B_720p, and pass it as --model-path. That directory
must contain config.yaml and checkpoints/SANA_Video_2.0_5B_720p.pth. Set
--component-paths.text_encoder /path/to/gemma-2-2b-it and
--component-paths.vae /path/to/ltx-2.3-vae. The latter directory may contain
the vae subdirectory or the VAE files directly. Without these overrides,
the component checkpoints are downloaded automatically.