Most tutorials tell you to use DPM++ 2M Karras and move on. That's not bad advice — it's a reliable default. But the sampler choice is not a neutral technical detail. It shapes the textural character of the output, determines how the model navigates the denoising trajectory, and interacts with your step count and CFG scale in ways that have significant quality consequences for atmospheric and cinematic image generation.
Understanding what samplers actually do at a practical level — not a mathematical one — lets you make deliberate choices rather than relying on defaults. This guide covers all major samplers available in AUTOMATIC1111 and ComfyUI, what each one does to output character and convergence behavior, and the specific configurations that produce best results for dark atmospheric and cinematic wallpaper generation.
What a Sampler Is Actually Doing
Stable Diffusion generates images by starting from pure noise and progressively removing that noise over a series of steps until a coherent image emerges. The sampler is the algorithm that calculates each noise-removal step — specifically, how it estimates the "correct" direction to move given the current noisy image state and the text conditioning signal.
Think of it like numerical integration: you're trying to trace a path from noise to image, and the sampler is the integration method that determines how you step along that path. Different numerical methods make different tradeoffs:
- Step efficiency: How much progress each step makes toward the final image. High step efficiency = fewer steps needed for convergence. Low step efficiency = needs more steps but each step may handle fine detail better.
- Stochasticity: Whether the sampler introduces controlled randomness into the trajectory. Deterministic samplers (same seed, same result always). Stochastic samplers (same seed, slightly different result each run due to randomness injected during sampling). Stochastic behavior creates more variation per seed — useful for exploration, less useful for reproducible production.
- Error accumulation: Some samplers make approximations that accumulate error over many steps. A sampler that approximates well at 20 steps may degrade at 40 steps. Others improve monotonically with step count.
- Convergence behavior: Whether the sampler converges to a stable final image as step count increases, or whether it continues to change at high step counts. Samplers that converge cleanly are better for knowing when you've reached quality ceiling.
The Noise Schedule: The Other Variable People Ignore
Before covering individual samplers, the noise schedule is often overlooked but critically important. The schedule defines how noise is distributed across the denoising steps — specifically, the sequence of noise levels from the initial high-noise state to the final low-noise state.
Uniform schedule: Equal step sizes across the full noise range. Simple but not optimal — many steps are "wasted" in the low-noise zone where changes are minimal.
Karras schedule: Non-uniform distribution that concentrates steps in the high-noise zone where they have the most impact on structural composition and content. Named after Tero Karras (the NVIDIA researcher who developed it). Adding "Karras" to a sampler name (DPM++ 2M Karras vs. plain DPM++ 2M) means it uses this schedule rather than a uniform one. Karras scheduling is almost always an improvement for photographic and atmospheric content — it makes early structural decisions more carefully.
Exponential and SGM schedules: Available in newer versions of A1111 and ComfyUI. Exponential is similar to Karras with different concentration. SGM (from the SDXL research paper) is the default schedule recommended for SDXL. Using a Karras sampler with SGM scheduling on SDXL often outperforms it with standard Karras.
In ComfyUI, the sampler and scheduler are separate nodes — you can combine any sampler with any scheduler. In A1111, some samplers have scheduler variants baked in (DPM++ 2M Karras has Karras built in), others use the global scheduler setting.
The Primary Samplers: What They Actually Do to Your Output
DPM++ 2M Karras
The current consensus default, and the default for good reason. DPM++ 2M uses a second-order multistep method — it considers both the current and previous denoising steps when estimating the next step direction. This multi-step lookback significantly improves convergence efficiency compared to single-step methods. Combined with Karras scheduling, it produces strong structural coherence and good mid-frequency detail at relatively low step counts.
For atmospheric night scenes and cinematic wallpapers, DPM++ 2M Karras handles the mid-tone gradation well — the atmospheric glow zones, soft light falloff, and depth haze that define this aesthetic. These mid-frequency details depend on the sampler's ability to make careful decisions in the mid-noise zone, and 2M Karras does this better than most single-step alternatives.
Recommended settings for dark atmospheric content:
- Steps: 22–28 (20 minimum, diminishing returns above 32)
- CFG scale: 6.5–7.5 (above 8 tends to over-saturate and lose atmospheric subtlety)
- Best for: final quality renders, consistent production runs
- Weakness: slightly "clean" character — can lack organic texture in foliage, fabric, and skin detail compared to SDE variants
DPM++ 2M SDE Karras
The SDE (Stochastic Differential Equation) variant of 2M Karras. The SDE component introduces controlled stochasticity into the denoising trajectory — each step includes a small random perturbation. This has two consequences:
First: outputs have a different fine-detail character than deterministic 2M Karras. The stochastic perturbations introduce more organic micro-texture, especially in areas where the model would otherwise produce slightly over-smooth results: atmospheric haze zones, cloud formations, rain effects, fabric, grass, loose organic material. For cinematic atmospheric content, this often produces more compelling results than deterministic sampling.
Second: the same seed produces slightly different results each time, because the stochastic component varies across runs. This means SDE Karras is less suitable for strict production reproducibility — but for exploration (generating many variations to find the best one), the additional variation is a feature rather than a bug.
Recommended settings:
- Steps: 28–35 (needs slightly more steps than deterministic 2M to reach equivalent structural coherence)
- CFG scale: 6–7 (slightly lower than 2M Karras to compensate for the stochastic energy)
- Best for: atmospheric weather effects, organic textures, exploration runs
- Weakness: slower convergence; more steps needed; less predictable for production
Euler and Euler a
Euler is one of the simplest samplers — a first-order single-step method named after the numerical integration method it implements. It's fast and predictable but not efficient — it needs more steps than multistep methods to reach equivalent structural quality.
Euler a (Euler Ancestral) adds an ancestral sampling component — a form of stochasticity similar to SDE but with different character. Euler a is highly creative and produces strong color energy at low step counts, but doesn't converge cleanly with increasing steps — quality peaks somewhere in the 20–25 step range and then starts changing character rather than improving. For atmosphericwallpaper use:
- Useful for exploration phases — generating many diverse candidates quickly to find compositional directions worth developing
- At 15–20 steps, produces impressionistic, painterly outputs with strong color character. This "impressionistic" quality works well for cozy interior scenes, fantasy environments, and heavily stylized content
- Less useful for precise architectural detail, crisp night city scenes where structural lines need to be clean
- The non-convergence property means: don't increase step count past ~25 expecting quality improvement. It will just change, not improve
Recommended settings:
- Steps: 18–25 (exploration), don't go higher expecting improvement
- CFG: 7–8.5 (tolerates slightly higher CFG than 2M Karras before over-saturation)
- Best for: exploration, stylized painterly content, high-volume draft generation
DDIM
DDIM (Denoising Diffusion Implicit Models) is an older deterministic sampler that was the standard before DPM++ variants became dominant. It requires significantly more steps than modern samplers to reach equivalent quality (40–50 steps vs. 22–28 for DPM++ 2M Karras) but has specific properties that make it valuable for certain workflows:
Absolute determinism: DDIM is fully deterministic. The same seed, same settings, same prompt will produce a pixel-identical result every time. No stochastic components anywhere in the process. This makes it the most reliable sampler for reproducible workflows where exact consistency matters.
Smooth latent-space interpolation: DDIM has smooth, predictable behavior in the latent space of the model. This makes it the best sampler for img2img workflows where you want controlled, predictable structural preservation at a given denoising strength. DPM++ variants can produce surprising structural changes at specific denoising values; DDIM's structural changes are monotonically proportional to denoising strength in a more predictable way.
For the img2img refinement workflow, DDIM at 40 steps is the reliable choice for Stage 3 and 4 refinement where you've committed to a specific structural composition and want predictable behavior during detail enhancement and upscaling.
Recommended settings:
- Steps: 40–50 for quality output (not worth using at lower step counts)
- CFG: 7–8
- Best for: img2img workflows, upscaling passes, production reproducibility
- Not for: fast iteration — the step count cost makes it slow
UniPC
UniPC (Unified Predictor-Corrector) is a newer high-efficiency sampler that uses a predictor-corrector structure to achieve excellent quality at very low step counts. 15–18 steps in UniPC often produces results equivalent to 25–28 steps in DPM++ 2M Karras.
The quality ceiling is slightly lower than DPM++ 2M Karras at matched step counts, but for atmospheric scenes without extreme fine-detail requirements (dense foliage, complex fabric, intricate architecture) the difference is minimal. For high-volume production where generation speed matters, UniPC at 18–22 steps is the most effective cost-quality-speed tradeoff.
Recommended settings:
- Steps: 18–22
- CFG: 6.5–7.5
- Best for: high-volume production runs, draft generation where speed matters
DPM++ 3M SDE Karras (newer A1111 versions)
The third-order multistep SDE variant. Uses three previous steps rather than two for the lookback. This produces higher quality at a given step count compared to 2M SDE, but the improvement is diminishing for most content types — most of the quality gain from multistep lookahead is captured by the second step, and the third adds relatively little. Slower per step and requires slightly more steps to stabilize. Worth testing for very high-resolution final outputs but not necessary for standard production.
