---

name: denoising-diffusion-pytorch description: "Use denoising-diffusion-pytorch for PyTorch DDPM/DDIM image diffusion, 1D sequence diffusion, conditioning and guidance, and advanced diffusion variants." disable-model-invocation: true metadata: disco-role: operating license: MIT

denoising-diffusion-pytorch

Use this repo skill when a Researcher needs to operate the public denoising-diffusion-pytorch package (denoising_diffusion_pytorch import) from self-contained guidance. It covers the inspected public API surface at version 2.3.1.

Install and sanity check

python -m pip install denoising-diffusion-pytorch
python - <<'PYCODE'
from denoising_diffusion_pytorch import Unet, GaussianDiffusion, Unet1D, GaussianDiffusion1D
print('denoising_diffusion_pytorch imports ok')
PYCODE

For a no-training environment check, run scripts/check_env.py. It verifies package metadata, public imports, a tiny CPU tensor operation, and optional CUDA visibility.

Route by task

User task	Read
2D image DDPM/DDIM, Unet, GaussianDiffusion, image folder Trainer, FID, RePaint/inpainting, image tensor shape errors	sub-skills/image-diffusion/SKILL.md
1D/time-series/sequence tensors, Unet1D, GaussianDiffusion1D, Dataset1D, Trainer1D, channel-first vs channel-last layout	sub-skills/sequence-diffusion/SKILL.md
classifier-free guidance, class labels, cond_scale, CFG++, external classifier cond_fn, XMWrapper multi-candidate loss	sub-skills/conditioning-guidance/SKILL.md
Karras UNets, video-shaped Karras 3D, continuous-time / v-param / EDM / simple diffusion, learned variance, weighted objective, flash/SDPA compatibility	sub-skills/advanced-variants/SKILL.md

Shared references

• references/package-overview.md maps public exports, dependencies, and sub-skill ownership.
• references/troubleshooting.md covers cross-cutting install/import, PyTorch/CUDA, Accelerate, optional FID, and smoke-test failures.
• references/repo-provenance.md records the source snapshot and refresh baseline.

Operating rules

• Use the bundled scripts for smoke tests; do not run long training or FID for routine verification.
• Keep data tensors normalized to [0, 1] unless auto_normalize=False is explicitly chosen.
• CPU is the default verification backend; CUDA, flash attention, FID, and Accelerate multi-GPU are optional accelerators unless a user explicitly asks for them.
• The package has no public CLI entry points. Workflows are Python API workflows.
• If a task asks to edit this repository rather than use the package, this repo skill can explain APIs, but repository-maintenance policy is outside the selected operating graph.