strands-diffusers

The universal entrypoint to HuggingFace diffusers for Strands agents. One tool — use_diffusers — wraps the whole library with zero hardcoding: discover and run any of its 300+ pipelines across every modality. It's a visual library, so here's what it actually produces — every asset below is real model output, not a placeholder:

text → image any of 108 image pipelines

text → video LTX · Wan · CogVideoX · Hunyuan

robot actions 🤖 Cosmos WFM: world video + actions

text → audio StableAudio · AudioLDM2

text / image / video / robot-state  IN
image / video / audio / actions / 3d  OUT

The registry is built at runtime from diffusers._import_structure, so new pipelines are supported automatically with no code change. Same philosophy as use_aws, use_lerobot, and use_transformers: discover, don't hardcode.

3D mesh ShapE - verts/faces to .ply

audio (hear the .wav) StableAudio - waveform to .wav

100% coverage, zero hardcoding

Every pipeline, model, and scheduler diffusers ships is reachable through one tool. When diffusers adds a new pipeline, use_diffusers exposes it immediately.

Physical-AI: world-foundation models with action outputs

"Put the pot to the left of the purple item."

"Pick up the cloth and place it in the bowl."

"Open the drawer and place the spoon inside."

Same robot, same first observation — different task prompt → different imagined world and different predicted actions. Five real rollouts + all three Cosmos action modes in the WFM gallery.

This is the headline. A Cosmos action-policy rollout predicts both a future world video and the robot action chunk that produces it. One use_diffusers(action="run", ...) returns a .mp4 world video, a .json action chunk (normalized [-1, 1], shape [num_chunks, T, action_dim]), and optional .wav sound — and you can see the motion:

time-series (every dim, gripper highlighted)

end-effector path (dims 0–2)

Verified end-to-end on NVIDIA Thor (nvidia/Cosmos3-Nano, bf16/cuda): one call produced a world video (17, 480, 640, 3) and an action chunk (1, 16, 10). See examples/cosmos_action_policy.py.

Install

pip install -e .
pip install -e ".[video,audio]"   # mp4 export, wav I/O

Quick start

from strands import Agent
from strands_diffusers import use_diffusers

agent = Agent(tools=[use_diffusers])
agent("Generate an image of a robot arm in a kitchen")
agent("Run a Cosmos action-policy rollout on robot.mp4 and give me the actions")

Direct:

use_diffusers(action="run", pipeline="StableDiffusionPipeline",
              model="stabilityai/stable-diffusion-2-1",
              parameters={"prompt": "a robot arm in a kitchen"})
# -> {"artifacts": ["/tmp/strands_diffusers/image_*.png"]}

Two layers

run loads a pipeline via from_pretrained and calls it; inputs are coerced (path / URL / base64 to PIL / video), outputs auto-saved and returned by path.

call resolves and calls any diffusers class, function, or method (schedulers, VAEs, CosmosActionCondition, utils). cached:key references resolve to live objects; "**" unpacks a cached mapping into kwargs.

use_diffusers(action="call", target="CosmosActionCondition",
              parameters={"mode": "policy", "video": "robot.mp4"}, cache_key="cond")
use_diffusers(action="run", pipeline="Cosmos3OmniPipeline", model="nvidia/Cosmos3-Nano",
              parameters={"prompt": "...", "action": "cached:cond"},
              dtype="bfloat16", device="cuda")

Discovery

Architecture

core/registry.py  zero-hardcode taxonomy from diffusers._import_structure
core/engine.py    load/cache pipelines, auto device+dtype
core/io.py        coerce inputs; serialize video/image/audio/action/mesh
core/viz.py       render robot action chunks to plots + animation
tools/use_diffusers.py  the single @tool: run + call + discovery

Testing

pip install -e ".[video,audio,dev]"
pytest tests/ -q          # unit tests, no GPU, no downloads
python examples/smoke.py  # E2E gate on tiny fixtures

Every visual in this README and the docs is produced by real use_diffusers calls — regenerate them with:

python examples/generate_docs_assets.py

Docs

📖 cagataycali.github.io/strands-diffusers — quickstart, full gallery (images / video / audio / actions / 3D), the world-foundation-model story, discovery, and the two-layer design.

MIT

MCP Server (Claude Code / Claude Desktop / Cursor / any MCP client)

pip install "strands-diffusers[mcp]"

strands-diffusers-mcp                    # stdio (Claude Code / Desktop)
strands-diffusers-mcp --http --port 8020 # HTTP multi-client

Claude Code: claude mcp add diffusers -- strands-diffusers-mcp

Claude Desktop:

{"mcpServers": {"diffusers": {"command": "strands-diffusers-mcp"}}}