Best Stable Diffusion Samplers & Settings for Uncensored Generation in 2026
Confused by DPM++, Euler, UniPC, CFG scale, and step counts? Here's what every Stable Diffusion setting actually does and the optimal configurations for SDXL, Flux, and SD 1.5 on local hardware.
Settings Are the Hidden Lever
Same prompt, same model, different settings — completely different output. The sampler, CFG scale, and step count determine how the image forms. Most people leave defaults and get mediocre results.
Here's the no-BS guide to what each setting does and the optimal values for every model architecture.
Sampler Cheat Sheet
| Sampler | Steps | Speed | Notes |
|---|---|---|---|
| DPM++ 2M Karras | 20–30 | Fast | Best all-around. Default choice for SDXL. |
| DPM++ SDE Karras | 15–25 | Fast | Slightly more creative/varied. Good alternative. |
| Euler a | 20–40 | Fast | Classic. Softer output. Great for anime models. |
| Euler | 20–40 | Fast | Deterministic version of Euler a. Converges cleanly. |
| UniPC | 15–25 | Very fast | Fewer steps needed. Good for low VRAM speed runs. |
| DDIM | 30–50 | Medium | Legacy. Needed for some ControlNet workflows. |
| DPM++ 3M SDE Karras | 20–30 | Slow | Highest detail at cost of speed. Good for final renders. |
TL;DR recommendation:
Use DPM++ 2M Karras at 25 steps for SDXL. You're done. Swap to Euler a if you're running anime models.
CFG Scale Explained
CFG (Classifier-Free Guidance) controls how strictly the model follows your prompt:
- CFG 1–3: Very creative, ignores prompt, often abstract
- CFG 4–6: Balanced creativity. Good for SDXL where models are already well-trained
- CFG 7–9: Standard range. Follows prompt closely. Most SD 1.5 models work best here
- CFG 10–15: Very strict. Can cause oversaturation and artifacts. Only use if output is too random
- CFG 15+: Usually broken output. Avoid unless a specific model recommends it
Per-model defaults:
- SDXL (Juggernaut, DreamShaper): CFG 5–7
- SD 1.5 (CyberRealistic, Deliberate): CFG 7–9
- Pony Diffusion: CFG 6–8
- Flux: Uses "guidance scale" not CFG — typically 3–4
Steps: More Isn't Always Better
Each step refines the image. But there are diminishing returns:
- 10 steps: Rough, often blurry — only useful for Turbo/Lightning models
- 20 steps: Good quality for most samplers. This is the starting point.
- 25–30 steps: Optimal for most use cases. Nearly converged output.
- 40+ steps: Minimal improvement. Wastes time. Only try if you see improvement between 30→40 with your specific model.
Exception — Turbo/Lightning: These distilled models are designed for 4–8 steps. Using 20+ steps on them produces worse results, not better.
How many hours have you spent tweaking settings? Every model has different optimal values. Every Reddit thread recommends something different. LocalForge AI gives you a private offline setup with tested defaults for every included model, so the settings in this guide are already configured. You can always customize, but you start with settings that work.
Other Settings That Matter
- Resolution: Generate at the model's native res (1024×1024 for SDXL, 512×512 for SD 1.5). Non-square ratios work but keep total pixel count similar.
- Seed: Set a fixed seed for reproducibility. Use -1 for random. If you like an image, save the seed — you can iterate on the prompt while keeping the composition.
- Clip Skip: For SD 1.5 anime models, set Clip Skip to 2. For photorealistic SDXL, leave at 1 (default).
- VAE: SDXL models usually bundle a VAE. If colors look washed out, explicitly load the SDXL VAE (sdxl_vae.safetensors).
- Hires Fix: Generates at a lower res, then upscales with img2img. Good for getting detail at higher resolutions but doubles generation time.
FAQ
Why does the same prompt give different results each time?
Random seed. Each seed produces a different starting noise pattern. Fix the seed to get reproducible results. In Forge: set Seed to a specific number instead of -1.
Do these settings affect uncensored content?
No. Samplers and settings are about image quality, not content filtering. On local uncensored models, there are no filters regardless of settings.
Should I use different settings for each model?
Ideally, yes. Model pages on CivitAI usually list recommended settings. As a starting point: DPM++ 2M Karras, 25 steps, CFG 7 works for almost everything.