---

name: sequence-diffusion description: "Use denoising-diffusion-pytorch for 1D sequence diffusion with Unet1D, GaussianDiffusion1D, Dataset1D, Trainer1D, tensor layouts, sampling, interpolation, and safe smoke checks." disable-model-invocation: true metadata: disco-role: operating license: MIT

Sequence Diffusion

Use this sub-skill when the task is about 1D diffusion over tensors such as audio-like features, time series, token embeddings, or other sequences using the public denoising_diffusion_pytorch API from denoising-diffusion-pytorch 2.3.1.

Read first

• For signatures, tensor layouts, supported objectives and schedules, read references/api-reference.md.
• For loss, sampling, interpolation, channel-last adaptation, and Trainer1D recipes, read references/workflows.md.
• For common assertion and runtime failures, read references/troubleshooting.md.
• To check an installed package without training, run scripts/smoke_sequence_diffusion.py.

Route here when

• The user wants Unet1D, GaussianDiffusion1D, Dataset1D, or Trainer1D.
• Input data is shaped as (batch, channels, sequence_length) or needs help with (batch, sequence_length, channels) conversion.
• The task mentions seq_length, sequence channels/features, 1D sampling, 1D interpolation, channel_first, sampling_timesteps, or Accelerate-backed Trainer1D basics.

Do not handle here

• 2D image folders, image Trainer, FID, or RePaint: route to ../image-diffusion/SKILL.md.
• XMWrapper, classifier-free guidance, or classifier-gradient guidance layered around a sequence model: route guidance setup to ../conditioning-guidance/SKILL.md, then return here for base 1D tensor layout.
• KarrasUnet1D or advanced diffusion objectives beyond standard GaussianDiffusion1D: route to ../advanced-variants/SKILL.md.

Minimal safe pattern

import torch
from denoising_diffusion_pytorch import Unet1D, GaussianDiffusion1D

model = Unet1D(dim=8, dim_mults=(1,), channels=2)
diffusion = GaussianDiffusion1D(model, seq_length=8, timesteps=8,
                                sampling_timesteps=4, objective='pred_noise',
                                beta_schedule='cosine')
x = torch.rand(2, 2, 8)   # (batch, channels, seq_length), values in [0, 1]
loss = diffusion(x)
sample = diffusion.sample(batch_size=2)
assert sample.shape == (2, 2, 8)