Skip to content

Schedulers

Introduction

In the context of diffusion models, a scheduler (also known as a sampler) is the algorithm responsible for managing the iterative process of turning pure noise into a clean, coherent image.

The Core Concept

While the model provides the "intelligence" to recognize patterns in noise, the scheduler provides the "roadmap" for how to navigate from chaos to clarity.

Diffusion models are trained by gradually adding noise to an image until it becomes unrecognizable. To generate a new image, the model performs this process in reverse. However, the neural network doesn't "jump" to the final image in one step. Instead, it predicts the noise present in its current input, and the scheduler calculates how to subtract that noise to move one step closer to the final result.

Speed vs Quality Trade-offs

The choice of scheduler significantly impacts the final result:

  • Efficiency: Schedulers like UniPC, LCM, or TCD are optimized for speed, producing high-quality images in just 4–8 steps
  • Precision: Schedulers like DPM++ 2M or Heun provide high mathematical accuracy and smoother gradients, typically requiring 20–30 steps for optimal results
  • Consistency: Deterministic schedulers (ODE-based) will always produce the same image for the same seed, while stochastic schedulers (SDE-based) may vary slightly even with the same seed

Model Compatibility

For a scheduler to work correctly, it must match the prediction type the model was trained with:

  • Epsilon: The standard for most models (e.g., SD 1.5, SDXL). The model predicts the noise added to the image
  • V-Prediction: Used in models like SD 2.1-v and certain SDXL refinements. The model predicts a velocity vector that balances noise and signal, which helps avoid color shifts at very high/low noise levels
  • Flow Matching: The modern standard for models like Flux and SD3. The model predicts a constant-speed linear path ("flow") between pure noise and pure data. This typically results in much faster convergence and higher detail retention

Important

Using mismatched scheduler/precition-type will result in pure noise or artifacts

List

List of schedulers available in SD.Next is broken down into 3 groups:

Diffusers

1. Foundational Gaussian Schedulers

The original samplers that defined the diffusion era, focusing on iterative denoising through Markovian and non-Markovian processes. - Foundational Math: Based on the original Ho et al. and Song et al. formulations. - Versatility: Support for inversion and parallel sampling.

Scheduler Description Variants
DDPMScheduler Denoising Diffusion Probabilistic Models; the standard for training and original inference. 2 (PT, Parallel)
DDIMScheduler Denoising Diffusion Implicit Models; enables faster, non-Markovian sampling and inversion. 4 (PT, Parallel, CogVideoX, Inverse)
PNDMScheduler Pseudo Numerical Methods for Diffusion Models; uses multi-step Runge-Kutta logic. 1

2. DPM-Solver Family

A suite of high-order solvers specifically designed to solve the Probability Flow ODE of diffusion models with fewer steps. - ODE Efficiency: Purpose-built for the semi-linear structure of diffusion ODEs. - Karras Support: Deep integration with Karras-style noise schedules.

Scheduler Description Variants
DPMSolverMultistepScheduler Efficient high-order multistep solver; the standard recommendation for Stable Diffusion. 4 (PT, Inverse, CogVideoX, Cosine)
DPMSolverSinglestepScheduler Higher-order singlestep solver for precise but slightly slower sampling. 1
DPMSolverSDEScheduler Stochastic variant using SDE integration for improved texture and variety. 1
EDMDPMSolverMultistepScheduler DPM-Solver implementation optimized for the EDM (Elucidating Design Space) framework. 1

3. Euler & Heun (Karras-style) Schedulers

Classical numerical methods adapted for discrete-time diffusion, often preferred for their predictable and clean convergence. - Prediction Modes: Heavy support for epsilon, v_prediction, and sample targets. - Ancestral Sampling: Includes ancestral variants that add noise at each step.

Scheduler Description Variants
EulerDiscreteScheduler Simple and effective first-order ODE solver; excellent for most SD models. 2 (PT, Flax)
EulerAncestralDiscreteScheduler Euler method with added stochastic noise; known for creative diversity and stability. 2 (PT, LTX)
HeunDiscreteScheduler Second-order method providing better accuracy than Euler at the cost of double model calls. 1
LMSDiscreteScheduler Linear Multi-step solver using a history of gradients to improve convergence. 2 (PT, Flax)

4. Modern Distillation & High-Speed Solvers

State-of-the-art solvers designed for 1-4 step inference through consistency or distillation techniques. - Extreme Speed: Enables near real-time generation. - Consistency Models: Based on the Consistency Models (CM) and Latent Consistency (LCM) research.

Scheduler Description Variants
LCMScheduler Latent Consistency Models; the industry standard for 4-step real-time generation. 2 (PT, FlowMatch)
TCDScheduler Trajectory Consistency Distillation; improves on LCM for better 1-step quality. 1
SCMScheduler Simple Consistency Models; streamlined distillation for high-quality fast sampling. 1
ConsistencyDecoderScheduler Specialized scheduler for the DALL-E 3 consistency decoder models. 1

5. Flow Matching & Rectified Flow

Schedulers for the latest generation of "Flow" models (like Flux, SD3, and AuraFlow) which use linear velocity targets. - Velocity Prediction: Designed specifically for models trained on flow-matching objectives. - Linear Trajectories: Optimal for models that denoise along a straight line.

Scheduler Description Variants
FlowMatchEulerDiscreteScheduler The standard Euler solver for Flow Matching models (SD3, Flux). 1
FlowMatchHeunDiscreteScheduler Higher-order Heun solver for Flow Matching trajectories. 1

SDNext

1. High-Precision & Advanced ODE Solvers

Solvers designed for superior convergence and precision, often using predictor-corrector or boundary-diffusion frameworks.

Scheduler Description Variants
BDIA_DDIMScheduler Boundary-Diffusion Implicit Algorithm; high-precision extension of DDIM with non-Markovian guidance and gamma-weighted trajectory correction. 2
DCSolverMultistepScheduler Dynamic Compensation solver; multi-step UniPC framework with dynamic extrapolation and ratio optimization to minimize approximation error. 2
TDDScheduler Time-Dependent Diffusion; experimental sampler that extends DPMSolverSinglestep with special jump logic and TDD-specific training step support. 3

2. Flow Matching Optimized Solvers

Specialized solvers designed for the latest generation of Flow-based models (e.g., Flux, SD3, AuraFlow), supporting resolution-aware trajectory shifting.

Scheduler Description Variants
FlowMatchDPMSolverMultistepScheduler Dedicated high-order DPM solver for Flow Matching with integrated Brownian Tree smoke/noise stability and multiple SDE/ODE modes. 7
FlowUniPCMultistepScheduler Multi-step UniPC framework adapted for flow-prediction models, supporting dynamic shifting and high-order B(h) updates. 1
FlashFlowMatchEulerDiscreteScheduler Optimized Euler-based scheduler for FlashFlow models, featuring resolution-aware dynamic shifting (mu/base/max shift). 1

3. Fast-Step & Distillation Solvers

Production-grade solvers optimized for extremely low step counts (1-4 steps) while maintaining visual fidelity.

Scheduler Description Variants
UFOGenScheduler Diffusion GAN-based sampler implementing both one-step and multi-step sampling trajectories with thresholding support. 1

4. Continuous & Variational Frameworks

Schedulers based on specific mathematical foundations for variational objectives and continuous time formulations.

Scheduler Description Variants
VDMScheduler Variational Diffusion Models; supports both discrete and continuous formulations of VDM objectives (linear, cosine, or sigmoid schedules). 1

Unique Features

  • Dynamic Compensation (DC): Implements dynamic extrapolation to correct for trajectory drift during the denoising process.
  • Brownian Tree Noise Sampler: Provides significantly more stable convergence in Flow Matching compared to standard random noise.
  • Resolution-Aware Shifting: Automatically adjusts noise schedules based on the image sequence length (total pixels) to optimize quality across resolutions.
  • BDIA Guidance: Uses boundary-diffusion implicit algorithm logic to guide DDIM samples with a configurable gamma-weighted correction.

RES4LYF

1. RES Family (Refined Exponential Solvers)

The core of the suite, implementing state-of-the-art exponential integration with high-order accuracy and perfect variance tracking. - High-Order Convergence: Maintains structural integrity at low step counts. - Variance Preservation: Eliminates brightness drift and color shift during generation. - Unified Interface: Seamless switching between multistep and multistage modes.

Scheduler Description Variants
RESUnifiedScheduler Unified interface for switching between various RES and DEIS integration schemes. 9
RESMultistepScheduler High-order multistep solver using historical gradients for efficient sampling. 4
RESMultistepSDEScheduler Stochastic multistep solver that maintains variance while improving sample diversity. 2
RESSinglestepScheduler Multi-stage singlestep solver offering high accuracy within a single timestep interval. 4
RESSinglestepSDEScheduler Stochastic variant of the singlestep solver for high-quality, diverse image generation. 4

2. Exponential Time Differencing (ETD) & Lawson

Advanced integrators that solve the linear part of the Probability Flow ODE exactly, providing superior stability for high-order updates. - Exact ODE Handling: Solves the deterministic part of the diffusion process without approximation. - Superior Stability: Prevents numerical explosions in high-order (3rd and 4th) sampling steps.

Scheduler Description Variants
ETDRKScheduler Implements Exponential Time Differencing Runge-Kutta methods for exact linear ODE handling. 5
LawsonScheduler Uses Lawson's transformation to simplify and stabilize exponential integration steps. 3
ABNorsettScheduler Implements Adams-Bashforth Norsett methods for stable multistep diffusion sampling. 3
DEISMultistepScheduler Diffusion Exponential Integrator Sampler utilizing multistep polynomial extrapolation. 3

3. Classical Numerical Integrators

Standard mathematical integrators optimized and refactored for the specific dynamics of the diffusion reverse process. - Familiar Tableaus: Uses proven RK, Radau, and Lobatto logic. - Modern Refactor: Fully updated to operate in normalized signal space for VP/VE compatibility.

Scheduler Description Variants
LinearRKScheduler Flexible implementation of standard Runge-Kutta methods from Euler to RK4. 7
RungeKuttaScheduler Classic nth-order, m-stage Runge-Kutta integrator with optimal balance. 3
SpecializedRKScheduler Collection of advanced solvers including SSPRK (Strong Stability Preserving) and TSI schemes. 5
LobattoScheduler High-order solvers based on Lobatto IIIA schemes, known for their strong stability properties. 3
RadauIIAScheduler Specialized solvers based on Radau IIA quadrature for robust stiff-ODE integration. 2
GaussLegendreScheduler High-precision symmetric solvers based on Gauss-Legendre quadrature. 3

4. Flow Matching & Physics-Inspired Samplers

Solvers designed for Flow Matching/Rectified Flow models and physics-based sampling dynamics. - Non-Euclidean Flows: Supports Hyperbolic, Spherical, and Lorentzian geometries. - Stochastic Refinement: Uses Langevin and tangent-based methods for unique textures.

Scheduler Description Variants
RiemannianFlowScheduler Solves flow matching problems on Euclidean, Spherical, Hyperbolic, and Lorentzian manifolds. 4
FlowEuclideanScheduler standard Euclidean metric flow matching variant. 1
FlowHyperbolicScheduler Flow matching on hyperbolic (Poincaré) manifolds. 1
FlowSphericalScheduler Flow matching on spherical manifolds. 1
FlowLorentzianScheduler Flow matching on Lorentzian manifolds. 1
PECScheduler Predictor-Corrector framework for refined, multi-pass sampling steps. 2
BongTangentScheduler Implements the Bong Tangent method for geometrically-guided diffusion sampling. 1
LangevinDynamicsScheduler Uses Langevin-style gradient steps for stochastic refinement of samples. 1
SimpleExponentialScheduler Lightweight solver using simple exponential decay for fast, low-step sampling. 1

5. Utilities & Sigma Generators

Infrastructure for controlling noise profiles and driving the integration process.

Scheduler Description Variants
CommonSigmaScheduler Centralized schedule generator for Arcsine, Easing, Sigmoid, and Sine noise profiles. 5
SigmaSigmoidScheduler S-shaped sigma profile for balanced noise distribution. 1
SigmaSineScheduler Sine-based profile for periodic noise modulation. 1
SigmaEasingScheduler Natural easing curves (In/Out/In-Out) for sigma progression. 1
SigmaArcsineScheduler Arcsine-based profile for concentrated sampling near boundaries. 1
SigmaSmoothScheduler Smoothstep-based profile for consistent noise transitions. 1

Credits

Based on the RES4LYF implementation for ComfyUI by @ClownsharkBatwing