Stable DiffusionParameters & settings

Stable Diffusion Settings Explained: Steps, CFG, Seed, Size

By Marco Cavazzana · Updated · 5 min read
ShareEmail
Short answer

Stable Diffusion’s core settings are the checkpoint (which model), sampler and steps (how the image is refined), CFG scale (how strictly the prompt is followed), seed (the random starting point), resolution (native 512 px for SD 1.5, 1024 px for SDXL and SD 3.5), and, for img2img, denoising strength. Sensible defaults are 20 to 30 steps, CFG around 5 to 7, native resolution and a fixed seed while you compare changes.

Stable Diffusion, from Stability AI, gives you more knobs than any hosted generator, and the same settings appear whether you run it locally in ComfyUI or the web UI, or on Stability’s platform and partner services. Once you understand what each one does, the knobs stop being intimidating and become the reason to use the model: you can hold everything constant and change one thing. This guide explains each setting, gives a starting value, and shows how they interact.

Checkpoint: which model you are running

The checkpoint is the model file itself. The main families are SD 1.5 (older, small, enormous community ecosystem), SDXL (higher resolution, better composition and text) and SD 3.5 (the most recent family, with large and medium variants). Within each family you will find fine-tuned checkpoints trained for particular looks such as photography, anime or architecture.

The checkpoint determines three things before you touch any other setting: the native resolution, which LoRAs are compatible, and the licence. Recent Stability models ship under a community licence with conditions for commercial use; read the model card for the exact terms, because they differ between versions.

Prompt and negative prompt

The positive prompt describes what you want; the negative prompt lists what to suppress. SD 1.5 and SDXL respond strongly to negatives ("blurry, deformed hands, text, watermark"). SD 3.5 and other newer models follow natural sentences better and need less negative prompting. See the Stable Diffusion prompts guide for structure and weighting syntax.

Sampler, scheduler and steps

Generation works by removing noise over a number of steps. The sampler (sometimes split into sampler and scheduler in ComfyUI) is the algorithm that does the removal, and steps is how many passes it takes.

  • Steps. More steps means more refinement up to a point, then nothing but time. 20 to 30 is the useful range for most samplers; some fast samplers and distilled models are designed for far fewer.
  • Sampler. Euler and DPM++ variants are the common defaults. "Ancestral" samplers (marked with an a) add noise each step, so they never fully converge and give a different image at 30 steps than at 20. Non-ancestral samplers converge, which is better for reproducibility.
  • Scheduler. Controls how noise is distributed across steps. Karras-style schedules often give cleaner detail at low step counts.

Pick one sampler and stay with it while you learn. Changing sampler and prompt together makes it impossible to know what caused a difference.

CFG scale: how hard the prompt pushes

Classifier-free guidance scale sets how strongly the model follows the prompt versus its own tendencies. Low values give soft, sometimes vague images that ignore details; high values give saturated, contrasty, over-sharpened results with artefacts. For SD 1.5 and SDXL, 5 to 7 is a good start. SD 3.5 generally wants lower values than older models. If colours look burnt or edges look crunchy, CFG is too high.

Seed: the starting noise

The seed sets the random noise the image grows from. Same seed, same settings, same prompt gives the same image, which is what makes comparison possible. Fix the seed while you adjust a prompt word or a setting; randomise it (often shown as -1) when you want variety. Always save the seed with a result you like.

Resolution: generate at native size

Each family was trained at a native size: around 512 px on the long side for SD 1.5, 1024 px for SDXL and SD 3.5. Generating far above native produces duplicated subjects and stretched anatomy; far below produces mush. Generate at native, then upscale with an upscaler or a hires pass. Aspect ratios are fine as long as the total pixel count stays close to native.

Denoising strength and LoRA weight

In img2img, denoising strength (0 to 1) sets how much of the source image is replaced. Around 0.3 keeps the picture and cleans it; 0.5 to 0.7 restyles while keeping composition; above 0.8 the source is a loose suggestion.

LoRAs are small add-on files that teach a style, character or object. Each has a weight, usually 0.6 to 1.0, and must match the checkpoint family. Stack two at most while learning; more than that and they fight.

Settings at a glance

SettingWhat it doesStart with
CheckpointThe model file; sets native size, LoRA compatibility and licenceSDXL or SD 3.5 for new work; SD 1.5 for its ecosystem
Negative promptThings to suppressblurry, deformed, text, watermark (less needed on SD 3.5)
Sampler / schedulerAlgorithm that removes noiseEuler or a DPM++ variant; Karras schedule
StepsNumber of refinement passes20 to 30
CFG scalePrompt adherence versus freedom5 to 7 for SD 1.5 and SDXL; lower for SD 3.5
SeedRandom starting noiseFixed while comparing, random for variety
ResolutionOutput sizeNative: ~512 px (SD 1.5), ~1024 px (SDXL, SD 3.5)
Denoising strength (img2img)How much of the source is replaced0.3 to clean, 0.5 to 0.7 to restyle
LoRA weightStrength of an add-on style or subject0.7, one or two LoRAs at a time

Interfaces label these slightly differently; ComfyUI exposes them as nodes, the web UI as sliders, and hosted services may hide some. The behaviour is the same underneath.

Try it

Checkpoint: an SDXL base or fine-tune
Prompt: A lone cyclist on a wet coastal road at dawn, low mist, pastel sky, wide shot, photographic, soft natural light
Negative: blurry, deformed, text, watermark
Sampler: DPM++ 2M, Karras   Steps: 25   CFG: 6
Seed: 12345 (fixed)   Size: 1216 x 832

Keep the seed and change only CFG to 3, then 10, to see the adherence curve for yourself. Then restore CFG 6 and change the sampler. One variable per run is the whole method.

Frequently asked questions

What CFG scale should I use in Stable Diffusion?
Start around 5 to 7 for SD 1.5 and SDXL. Lower values give the model more freedom and softer images, while higher values force prompt adherence but cause burnt colours and crunchy edges. SD 3.5 and other newer models generally want lower CFG than the older families.
How many steps do I need in Stable Diffusion?
Twenty to thirty steps is enough for most samplers, and beyond that you mostly add time. Fast or distilled models are designed for far fewer steps, so read the model card. Ancestral samplers keep changing the image with more steps, while non-ancestral ones converge and are better for reproducibility.
What resolution should I generate at in Stable Diffusion?
Generate at the model’s native size: roughly 512 px for SD 1.5 and 1024 px for SDXL and SD 3.5, keeping the total pixel count close to that for other aspect ratios. Going far larger duplicates subjects and stretches anatomy. Upscale afterwards with an upscaler or a hires pass.

Nothing you makeis lost.

Free · No card · Read-only · 2-minute setup

DreamdriveFree plan · Chrome
Add to Chrome — it’s free