Diffusers
Hugging Face's PyTorch library of pretrained diffusion pipelines for image, video and audio generation, for developers who run models from Python.
Apple Silicon support is via PyTorch MPS (docs: optimization/mps); the Stable Audio 3 docs warn that float16 on MPS produces noise.
At a glance
- Built by
- Hugging Face
- Status
- Active
- Latest release
- v0.41.0 (2026-10-06)
- License
- Apache-2.0
- Written in
- Python
- Serves
- Image, Video, Audio
- Hardware
- NVIDIA GPU, Apple Silicon (MPS), Intel Gaudi, AWS Inferentia/Trainium (via Optimum Neuron), Google TPU, CPU
- API
- Library only
- Quantization
- bitsandbytes, GGUF, torchao, Quanto, NVIDIA ModelOpt, Nunchaku, AutoRound, SDNQ
From Diffusers's own repo and docs, checked 21 Oct 2026.
Get started
pip install --upgrade diffusers[torch]DiffusionPipeline.from_pretrained("{model}")Replace {model} with the model ID below.
What it's good at
- Library, not a server
Pipelines are loaded and called from Python (from_pretrained, then pipe(prompt)); there is no built-in HTTP server. source ↗
- Modular Diffusers is now stable
v0.40.0 graduated Modular Diffusers from experimental; some new models (MiniMax-H3, MiniMax Music 3, Wan-Animate-2, FLUX 3 Image Klein) ship as modular pipelines only, loaded with ModularPipeline. source ↗
- Many quantization backends
Docs cover bitsandbytes, GGUF, torchao, Quanto, ModelOpt, Nunchaku, AutoRound and SDNQ (int8 down to 2-bit and FP8). source ↗
- Tensor parallelism
v0.40.0 added minimal model-agnostic tensor-parallel sharding, wired into the FLUX.1, FLUX.2 and Qwen-Image DiTs. source ↗
- New models often land on main first
Several model cards tell you to install from source (pip install git+https://github.com/huggingface/diffusers) until the next release ships the pipeline. source ↗
Catalog models it runs 12
DiffusionPipeline.from_pretrained("stabilityai/stable-audio-3-medium")StableAudio3Pipeline, but Stability's checkpoints are not published in Diffusers format; convert locally with scripts/convert_stable_audio_3_to_diffusers.py first.
DiffusionPipeline.from_pretrained("Qwen/Qwen-Image-2.1")QwenImage21Pipeline (generation, editing, RGBA output).
DiffusionPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")Modular pipeline only (ModularPipeline), on the main branch but not in v0.41.0; install from source.
DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.2-dev")Flux2Pipeline; BFL's model card says FLUX.2 [dev] is available in Diffusers.
DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B")Flux2KleinPipeline.
DiffusionPipeline.from_pretrained("ideogram-ai/ideogram-4-nf4")Ideogram4Pipeline; Ideogram's model zoo lists Diffusers support for the nf4 checkpoint (CUDA) but not the fp8 one.
DiffusionPipeline.from_pretrained("krea/Krea-2-Raw")Krea2Pipeline; the Krea-2-Turbo distilled checkpoint is also supported.
DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large")StableDiffusion3Pipeline; Stability's model card has a 'Using with Diffusers' section.
DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-3.5-medium")StableDiffusion3Pipeline; Stability's model card has a 'Using with Diffusers' section.
DiffusionPipeline.from_pretrained("diffusers/LTX-2.3-Diffusers")LTX2Pipeline with the converted diffusers/LTX-2.3-Diffusers repo; Lightricks' own LTX-2.3 card still says Diffusers support is 'coming soon'.
DiffusionPipeline.from_pretrained("Lightricks/LTX-2.5-Diffusers")Reuses LTX2Pipeline; Lightricks publishes the Diffusers-format repo.
DiffusionPipeline.from_pretrained("MiniMaxAI/MiniMax-H3")Modular pipeline only: ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3"); MiniMax's card points to the Diffusers docs.
Only models with an official source are listed. Diffusers may run others with the same architecture; see its supported-models page.