Scheduler

Also known as: scheduler, sampler

The scheduler (also called a sampler) determines the mathematical algorithm used to iteratively remove noise from the image. Different schedulers trade off speed, quality, and visual style — some converge faster while others produce smoother or more detailed results.

What It Does

During image generation, the diffusion model starts from pure random noise and needs a strategy for stepping toward a clean image. The scheduler defines exactly how noise is removed at each step — how much to remove, in what order, and with what mathematical corrections. Think of it as the roadmap the model follows through "noise space" on its way to a finished image. Two different schedulers given the same prompt and seed can produce noticeably different results, even with identical settings everywhere else.

Schedulers fall into a few families. The DPM++ family (DPM++ 2M, DPM++ SDE, and their Karras variants) use higher-order ODE solvers to converge quickly, often producing sharp results in fewer steps. Euler and Euler Ancestral are simpler first-order methods — Euler is deterministic while Euler Ancestral adds a small amount of noise at each step, which can introduce creative variation but makes results less reproducible. DDIM is an older but still popular option that is fully deterministic and allows smooth interpolation between images. Newer schedulers like UniPC aim for the best of both worlds: fast convergence with high quality.

The right scheduler depends on your model and how many steps you're using. Some schedulers need 30 or more steps to converge properly, while others look great at 20. Certain models ship with a recommended scheduler baked in, and the options you see in the dropdown are specific to each model — not every scheduler works with every architecture. When in doubt, start with the model's default and experiment from there.

Common Schedulers

DPM++ 2M Karras

Fast, high-quality, and the go-to default for SDXL-based models. Uses a second-order solver with a Karras noise schedule that front-loads detail, producing smooth, well-defined results in 20–30 steps.

DPM++ SDE Karras

Adds stochastic noise injection to DPM++. This introduces subtle variation that can enhance textures and organic details — skin, foliage, fabric — at the cost of slightly less reproducibility. Popular for photorealistic and portrait work.

Euler

The classic, general-purpose scheduler. Simple first-order solver that is deterministic and reliable across a wide range of step counts and models. A safe starting point if you're unsure which scheduler to pick.

Euler Ancestral (Euler A)

Like Euler but adds random noise at each step. This makes outputs non-deterministic — the same seed can produce different images — but the added randomness can yield more creative, painterly results with interesting variation.

DDIM

Denoising Diffusion Implicit Models. Fully deterministic and consistent, making it ideal when you need exact reproducibility. May need more steps than DPM++ to reach the same quality, but supports smooth interpolation between latents.

UniPC

Unified Predictor-Corrector. A newer scheduler designed for fast convergence with minimal artifacts. Can produce good results in as few as 10–15 steps, making it useful when you want to save generation time without losing too much quality.

PNDM / LMS

Older schedulers you may still encounter on some models. They work but are generally outperformed by DPM++ and UniPC. If you see them in a dropdown, they're fine to try but unlikely to be the best option available.

Visual Comparison

Image pending

scheduler = DPM++ 2M Karras

Image pending

scheduler = Euler

Image pending

scheduler = DDIM

Tips

DPM++ 2M Karras is a strong default for SDXL-based models — start there unless you have a reason not to.
If your image looks noisy even at 30+ steps, try switching schedulers before adding more steps. A better scheduler often fixes what extra steps cannot.
DDIM is your best bet when you need exact reproducibility across runs with the same seed.
Euler Ancestral can break you out of a rut — its randomness sometimes finds compositions that deterministic schedulers miss.
Scheduler and step count are linked: a fast scheduler like UniPC at 15 steps may outperform a slower one at 30. Experiment with both together.
Not all schedulers are available on all models — the options shown in the dropdown are specific to each model's architecture.