Implementation:AUTOMATIC1111 Stable diffusion webui UniPC Sampler

Knowledge Sources	AUTOMATIC1111_Stable_diffusion_webui
Domains	Sampling, Diffusion_Models
Last Updated	2025-05-15 00:00 GMT

Overview

Implements the UniPC (Unified Predictor-Corrector) sampling algorithm as a wrapper around the diffusion model for high-quality image generation with configurable step scheduling.

Description

The UniPC Sampler module provides the UniPCSampler class that wraps a diffusion model to perform denoising sampling using the UniPC algorithm. The sampler initializes by extracting the cumulative product of alphas from the model and constructs a discrete noise schedule. It supports both SD 1.X ("noise" parameterization) and SD 2.X ("v" parameterization) models. The sample method creates a classifier-free guidance model wrapper, initializes the UniPC solver with configurable variant, skip type, order, and lower-order-final options (all exposed through shared.opts), and runs the multistep sampling process. The sampler supports hook functions (before_sample, after_sample, after_update) that can be registered via set_hooks for external monitoring or modification of the sampling process. Conditioning can be provided as a dict, list, or tensor.

Usage

Use this sampler when the UniPC sampling method is selected in the WebUI. It provides high-quality results with fewer steps compared to some other samplers, supporting both SD 1.X and SD 2.X model architectures.

Code Reference

Source Location

• Repository: AUTOMATIC1111_Stable_diffusion_webui
• File: modules/models/diffusion/uni_pc/sampler.py
• Lines: 1-101

Signature

class UniPCSampler:
    def __init__(self, model, **kwargs) -> None
    def register_buffer(self, name: str, attr) -> None
    def set_hooks(self, before_sample, after_sample, after_update) -> None
    def sample(self, S, batch_size, shape, conditioning=None, callback=None,
               normals_sequence=None, img_callback=None, quantize_x0=False,
               eta=0., mask=None, x0=None, temperature=1., noise_dropout=0.,
               score_corrector=None, corrector_kwargs=None, verbose=True,
               x_T=None, log_every_t=100, unconditional_guidance_scale=1.,
               unconditional_conditioning=None, **kwargs) -> tuple[torch.Tensor, None]

Import

from modules.models.diffusion.uni_pc.sampler import UniPCSampler

I/O Contract

Inputs

Name	Type	Required	Description
model	object	Yes	The diffusion model with alphas_cumprod, parameterization, and apply_model attributes
S	int	Yes	Number of sampling steps
batch_size	int	Yes	Number of images to generate simultaneously
shape	tuple[int, int, int]	Yes	Latent space dimensions (C, H, W)
conditioning	dict, list, or Tensor	No	Positive conditioning embeddings
unconditional_conditioning	object	No	Negative/unconditional conditioning embeddings
unconditional_guidance_scale	float	No	Classifier-free guidance scale (default 1.0)
x_T	torch.Tensor	No	Initial noise tensor; generated randomly if not provided

Outputs

Name	Type	Description
samples	torch.Tensor	The denoised latent samples tensor
intermediates	None	Always None (placeholder for API compatibility)

Usage Examples

from modules.models.diffusion.uni_pc.sampler import UniPCSampler

sampler = UniPCSampler(model)

samples, _ = sampler.sample(
    S=20,
    batch_size=1,
    shape=(4, 64, 64),
    conditioning=positive_cond,
    unconditional_conditioning=negative_cond,
    unconditional_guidance_scale=7.5,
)

Related Pages

• Principle:AUTOMATIC1111_Stable_diffusion_webui_Sampling_Architecture