Different Sampler Components Breakdown for DiT models (Anima, Flux, Krea2, Klein, Qwen, Etc)

A diffusion model can be viewed as solving a differential equation backward; repeatedly estimating where the image should move next. A sampler is the numerical method used to make those estimates. Different methods have diff trade offs.

DPM / DPM++

• DPM — Diffusion Probabilistic Model Good: quality/efficiency, especially at less steps. Bad: more complicated. May introduce artifacts.
• ++ / DPM ++ Uses the DPM-Solver++ formulation. Good: stability and quality under CFG. Bad: doesn't automatically mean “better” than every other solver

1 / 2 / 3

• Order — 1st/2nd/3rd-order numerical approximation.
• Good: fewer-step sampling can benefit and be more accurate.
• Bad: higher order isn't automatically better. High step counts make similar images.

M = Multistep

• Reuses previous denoiser evaluations/history for order corrections.
• Good: strong efficiency. gets higher-order behavior without too much time re - evaluating.
• Bad: can be more sensitive to unusual timestep schedules/rapidly changing trajectories.

S = Single-step

• Does its higher-order correction within the current step.
• Good: useful when you don't want to depend as heavily on accumulated history.
• Bad: requires more computation per step

a = Ancestral

• Adds stochasticity/noise during the trajectory.
• Good: variation, softer/less rigid results, exploration.
• Bad: Can introduce instability/ make details less consistent/ OCs do not look same everytime.

SDE = Stochastic Differential Equations

• Uses a stochastic sampling formulation rather than purely deterministic trajectory integration.
• Good: more variation, robustness to certain trajectories. more natural results.
• Bad: less reproducible tractoriee, more randomness.

Heun

• A predictor/corrector integration style.
• Good: can improve trajectory accuracy and stability.
• Bad: more computation and time.

Euler

• Simple, first-order trajectory integration.
• Good: cheap, predictable, baseline. More universal.
• Bad: less accurate per step than other methods.

Dy = Dynamic-resolution modification

• Temporarily changes the sampling resolution to keep parts of denoising in a more comfortable scale for the model.
• Good: May improve large/unusual-resolution composition and anatomy on models.
• Bad: The creator found its benefits to be varied strongly with training resolution. (Mostly benefits SDXL)

SMEA = Sinusoidal Multipass Euler Ancestral

• Adds a multipass/resolution-oriented modification to Euler Ancestral.
• Good: designed to help structural coherence, limbs and hands, especially at large/unusual resolutions.
• Bad: costs more compute. (Mostly SDXL)

Negative

• A custom latent sign/manipulation technique.
• Good: can produce interesting alternative trajectories.
• Bad: its creator explicitly doesn't claim a theoretical quality advantage.(Mostly SDXL)

K

• K-Diffusion's implementation/family naming.
• Good: Less steps, higher details**.**
• Bad: none that I found.

UniPC

• Uni = Unified
• PC = Predictor-Corrector ** **Combines prediction and correction to get efficient high-order behavior. Good: useful to get good results with fewer model evaluations. Bad: advantage depends heavily on the model and schedule.

RES = Refined Exponential Solver

• An exponential-integrator family designed specifically for diffusion sampling.
• Good: efficient trajectory approximation. For low-step and high-efficiency sampling.
• Bad: less universally tested.

Adaptive

• Adjusts step sizes according to estimated numerical error instead of blindly using fixed steps.
• Good: spends computation where it is actually needed.
• Bad: step count may be less comparable to other fixed-step samplers.

Fast

• Speed-oriented DPM integration strategy.
• Good: efficient model evaluations.
• Bad: trades some accuracy and control for speed.

Flux2

• Indicates a variant adapted to the Flux.2 sampling regime/ modified for DiT models.
• Good: appropriate when working with that specific model/sampling formulation.
• Bad: Model specific. ER = Extended Reverse-Time: Good quality, low steps.

Note: Accuracy means accuracy to the solved DE! Not hand and other things! The model itself may know bad hands/ anatomy to be accurate!

Leave comment so I know ppl read this and I do more.