6 Improving Images with Samplers

Stable Diffusion generation is described as a process of starting from noise and gradually removing it until an image appears, and samplers are the tools that control how that removal happens. The chapter uses a haircut metaphor to explain that the sampler is like the stylist’s tool choice, while the schedule type is like deciding how much to trim at each step. Although these settings are less dramatic than changing checkpoints, they can still make the difference between a merely good image and one that feels polished and complete.

The chapter focuses on three samplers: DDIM, Euler, and DPM++ 2M. DDIM represents an earlier efficient approach, Euler reflects a classical ODE-solving method that became widely usable in diffusion workflows, and DPM++ 2M is a newer higher-order solver designed for faster convergence. By keeping the prompt, model, seed, and other settings fixed while changing only the sampler, the text shows that each method can produce subtly different results, and that no sampler is universally best; the “right” one depends on the image and the creator’s preferences.

Schedule types are shown to control how the amount of noise removed changes over the course of sampling. Uniform scheduling removes noise in equal increments, while exponential and Karras reduce it more gradually over time, often giving more refinement in later steps. The chapter also highlights ancestral samplers such as Euler a, which inject randomness at each step and may never fully converge to a single final image, making them useful when the goal is to explore controlled variations. Overall, the main lesson is that sampler and schedule choices are fine-tuning tools best used near the end of the image-making process.

We want to sculpt wild hair into something more beautiful

We want fast generation but not at the expense of good results.

A1111 over many choices for Sampling Method.

Each column is the same prompt, each row the same sampling method

There are many different schedule type options.

The only difference between these images is schedule type.

Exploring 10 steps with 3 different sampling methods.

Increasing the sample steps from 10 to 30.

Increasing the CFG scale shows limitations in higher order samplers.

At a higher step size DPM++ 2M no longer suffers from higher CFG.

Results of several different schedule types for the same configuration.

The variation between scheduler types is slightly more pronounced.

Ever evolving results of ancestral sampling

Cake found searching nearby images using an ancestral sampler.

• Sampler methods control how Stable Diffusion removes noise from an image.
• Schedule types control how much noise is removed at each step.
• Samplers provide subtle changes, but those subtle changes are key to getting an image “just right”.
• The CFG scale value can impact some samplers more than others at low step size.
• There is no universally best sampler, and often different samplers will be your subjective favorite for different prompts.
• Schedule types can provide even more subtle changes and should be tweaked at the end of your experiments with other configuration options.
• Ancestral samplers provide an additional method of injecting controlled randomness into your image generation process.

FAQ