Find any pipeline

You don't need to memorize class names. use_diffusers reads the installed diffusers library at runtime and tells you exactly what it can run. Add a new pipeline to diffusers and it shows up here on its own — nothing to update.

image · 108 video · 36 inpainting · 33 controlled-image · 27 image-to-image · 26 image-to-video · 18 other · 16 audio · 9 text-to-image · 9 video-to-video · 7 text-to-video · 6 upscaling · 6 video-to-world · 2 3d · 1

Ask what's available

action	returns
pipelines	all pipeline classes + derived modality
models	all model/backbone classes
schedulers	all scheduler classes
tasks	task → pipeline map
modalities	pipelines grouped by modality
wfm	world-foundation / action models
pipeline_info	constructor signature + docstring for one pipeline
inspect	signature + docs for any class, function, or method
cache / clear_cache	manage loaded pipelines in memory

Browse by modality

from strands_diffusers import use_diffusers

groups = use_diffusers(action="modalities")["data"]
# {'image': 108, 'inpainting': 33, 'video': 36, 'controlled-image': 27,
#  'image-to-image': 26, 'image-to-video': 18, 'other': 16, 'audio': 9,
#  'text-to-image': 9, 'video-to-video': 7, 'text-to-video': 6,
#  'upscaling': 6, 'video-to-world': 2, '3d': 1}

groups["audio"]    # ['AudioLDM2Pipeline', 'MusicLDMPipeline', 'StableAudioPipeline', ...]
groups["3d"]       # ['ShapEPipeline', ...]

Inspect before you run

Don't remember a pipeline's arguments? Ask:

use_diffusers(action="pipeline_info", target="StableDiffusionPipeline")
# -> constructor signature, accepted call kwargs, and the docstring

use_diffusers(action="inspect", target="DDIMScheduler")
use_diffusers(action="inspect", target="utils.export_to_video")

Why this matters

Hardcoded lists go stale the day diffusers ships a new model. Reading the library at runtime means:

• the list always matches your installed version,
• the agent can browse options and pick the right one, and
• there's nothing to maintain when new pipelines land upstream.

Same idea as use_aws, use_lerobot, and use_transformers: ask, don't hardcode.