strands-sana

NVIDIA Sana inference + training as Strands Agent tools.
33 tools · 15 models · 7 pipelines · MP4 video · LoRA · ControlNet · Sol-RL.

_{Sana-1.5 1.6B (20 steps) vs Sana-Sprint 1.6B (2 steps) — same prompt, same seed.
"a majestic rubber duck wearing aviator goggles, perched on a futuristic motorcycle in a cyberpunk neon-lit city, raining"}

_{sana_video_generate — 5s sunrise · 2s cyberpunk teaser, 480p, generated on NVIDIA Thor}

---

Why strands-sana?

pip install strands-sana

from strands import Agent
from strands_sana import sana_generate, sana_video_generate

agent = Agent(tools=[sana_generate, sana_video_generate])
agent("Generate a cyberpunk cityscape at 1024px, then animate it as a 5s clip")

That's it. The agent picks the right model, sets sane defaults, downloads weights on first use, and saves outputs. All 33 tools work the same way.

---

How it fits together

graph LR
    User([🗣️ Agent prompt]) --> Tools

    subgraph Tools[" 33 @tool functions "]
        T2I["🖼️ sana_generate<br/>sana_sprint_generate<br/>sana_inpaint<br/>sana_controlnet_generate"]
        IMG["🎨 sana_img2img"]
        VID["🎬 sana_video_generate<br/>sana_image_to_video"]
        TRAIN["🎓 sana_train_lora<br/>sana_train_solrl<br/>sana_train_video<br/>+4 more"]
        EXTRAS["⚙️ schedulers · quant · LoRA<br/>metrics · ComfyUI export<br/>HF upload · prefetch"]
    end

    Tools --> Wrapper["SanaPipelineWrapper<br/>(7 pipeline kinds)"]
    Wrapper --> Diffusers["🧨 diffusers ≥ 0.32"]
    Wrapper --> Upstream["📚 NVlabs/Sana<br/>(training scripts)"]
    Diffusers --> GPU([⚡ CUDA / MPS / CPU])

    style Tools fill:#5e35b1,stroke:#333,color:#fff
    style Wrapper fill:#7e57c2,stroke:#333,color:#fff
    style Diffusers fill:#ec407a,stroke:#333,color:#fff

Inference path uses diffusers directly — fast, well-maintained, GPU-accelerated. Training path shells out to upstream NVlabs/Sana scripts (auto-discovered).

---

What you can do

• 🖼️ Generate images

  Sana 1.0 / 1.5 / Sprint, plus PAG, Inpainting, ControlNet, multi-image batches.

  → Image generation

• 🎬 Generate videos

  SANA-Video 480p / 720p / LongSANA. Text-to-video and image-to-video.

  → Video generation

• ⚡ Sana-Sprint

  1-2 step distilled inference. ~0.1s per 1024px image on H100.

  → Sprint

• 🎨 Image-to-image

  Restyle existing images at Sprint speed.

  → Img2Img

• 🎓 Train your own

  LoRA, full pretrain, sCM-LADD distillation, Sol-RL post-training, video FSDP.

  → Training

• 🔬 Inference scaling

  Generate K candidates, pick best by CLIP / NVILA.

  → Inference scaling

• ⚙️ Memory modes

  Run 1.6B in 8 GB VRAM via int4/int8 quantization + offload.

  → Memory & quantization

• 🔌 Schedulers

  10 alias (flow-match-euler, DPM-Solver, Euler, DDIM, …) — swap at runtime.

  → Schedulers

• 📊 Metrics

  CLIPScore, ImageReward — score generations programmatically.

  → Metrics

---

Tested on real hardware

11 bugs were caught and fixed by stress-testing every tool on an NVIDIA Thor dev kit (CUDA 13). See the full bug log.

→ Continue to Quickstart or browse the Gallery.