Generation Parameters

This page explains the generation settings that matter most for beginners. If you are new to Stable Diffusion, the short version is:

• Prompt tells the model what you want to see.
• Negative Prompt tells the model what you want to avoid.
• Scheduler decides how the model walks from noise to an image.
• Steps decides how long that walk lasts.
• CFG decides how strongly the model listens to your prompt.
• Denoise Strength decides how much an input image can change in img2img.
• Seed decides which random starting noise is used.

You do not need to tune everything at once. Start with a familiar model, keep most settings near their defaults, and change one parameter at a time.

Just as important: there is no single best parameter set for every model. Different checkpoints are trained and merged differently, so they can prefer different schedulers, step counts, and CFG ranges.

TIP

CLIP caching: Local Dream persists text-encoder outputs to disk, so any prompt you have encoded before is reused on any later run — not just the most recent one. The cache is kept per model, because each model has its own text encoder; clearing one model's cache does not affect the others. You can clear caches at any time from Settings → File Manager.

What a cache hit saves (SDXL workflow, indicative):

• Skips the CLIP forward pass — about 1 s.
• Under Low RAM Mode, also skips loading and unloading the CLIP model — about 4 s.

Limitation: prompts that contain textual inversion embeddings are neither read from nor written to the cache — the text encoder runs every time for those.

Recommended Starting Points

If you just want a safe beginner setup:

• Prompt: keep it short and concrete
• Negative Prompt: only add a few obvious problems to avoid
• Scheduler: DPM++ 2M (Karras off is fine as a default)
• Steps: 20-30
• CFG: 5-8
• Denoise Strength: 0.75-0.85 for img2img
• Seed: any value, then keep it if you want to iterate on the same composition

These are safe starting points, not hard rules. If the model's official release page gives example settings, use those as your first reference and treat this page as the fallback.

Model-Specific Recommendations Matter

Many beginners assume parameters work like a universal preset. In practice, model authors often publish example images together with the settings they used, and those examples are usually the best place to start.

Why this matters:

• Some models were tuned around a specific scheduler
• Some models become harsh or oversaturated at high CFG
• Some models keep improving at higher steps, while others plateau early
• Fast workflows like LCM only make sense for models or setups that support them well

So the best workflow is usually:

1. Start from the model author's example settings.
2. Reproduce that general range inside Local Dream.
3. Only then make small adjustments for your own prompt or device.

If you are browsing the built-in models in this guide, use the Source links on Available Models to open each model's official page and check whether the author recommends a sampler, step count, CFG value, clip skip, trigger words, or negative prompt.

Prompt

The prompt, sometimes called the positive prompt, is the text that describes what you want to generate.

In general, prompts should be written in English. Stable Diffusion models are overwhelmingly trained and tagged around English prompt text, so English usually gives the most reliable and predictable results.

Examples:

• 1girl, garden, sunset, cinematic lighting
• a cozy cabin in the snow, warm window light, detailed, evening

For beginners, a good prompt is usually:

• Short
• Specific
• Written as a list of visual ideas

Good beginner habits:

• Write the prompt in English
• Start with the main subject first
• Add a few important details after that
• Only add style terms if you actually want that style

For example, this is easier to control:

1girl, white dress, standing in a garden, sunset, soft light

than a very long prompt stuffed with many styles, camera terms, and quality tags at once.

Non-English text may still work in some cases, especially with translation-assisted workflows or tag dictionaries, but it is usually less reliable than direct English prompting.

If the model's official page gives trigger words or example prompts, try those first. Some checkpoints are strongly tuned around a particular style tag, character tag, or naming convention.

Negative Prompt

The negative prompt tells the model what you want to avoid.

This can help reduce common problems such as:

• blurry details
• bad hands
• extra fingers
• extra limbs
• distorted anatomy
• low quality backgrounds

A negative prompt is not always required, and using a huge one is not automatically better.

For beginners, start simple. A small negative prompt like this is often enough:

low quality, blurry, bad anatomy, extra fingers

Important things to know:

• Too much negative prompting can fight against your main prompt
• Some models expect a specific negative prompt from the author
• Some models work fine with little or no negative prompt

If a model's official release page shows a recommended negative prompt, use that version first before inventing your own.

Scheduler

Available options:

• DPM++ 2M
• DPM++ 2M SDE
• Euler A
• Euler
• LCM

The scheduler controls the denoising path. In practice, it changes the style of the result, how stable the composition is, and how many steps are useful.

A separate Karras toggle sits next to the scheduler picker. When enabled, it switches the noise schedule to the Karras sigma curve, which usually pushes more steps into the low-noise region near the end of sampling. It tends to give slightly cleaner detail at moderate step counts. Karras is available for every scheduler except LCM.

DPM++ 2M

The best default choice for most users.

• Usually gives the most balanced results
• Works well for general txt2img and img2img
• Benefits from normal step counts like 20-30
• Pairs well with Karras when you want a little extra detail clarity

Use this when:

• You are unsure which scheduler to pick
• You want quality and stability
• You are following prompts from general SD guides

DPM++ 2M SDE

A stochastic (SDE) variant of DPM++ 2M. Each step adds a small amount of fresh noise, which tends to produce a slightly different — often more detailed or textured — look than the deterministic version.

• Good for richer textures and a less "polished" feel
• Usually works at similar step counts to DPM++ 2M (20-30)
• Often combined with Karras for sharper detail
• Because it injects noise, results at the same seed will differ from DPM++ 2M

Use this when:

• You want more texture and variation than plain DPM++ 2M
• A model author specifically recommends a DPM++ SDE sampler

Euler A

A classic ancestral sampler that often feels a bit more loose or creative.

• Can produce lively textures and stylized results
• Sometimes less stable than DPM++ 2M
• Usually works best with similar step counts, around 20-30

Use this when:

• You want to try a different feel without changing the prompt much
• A model page specifically recommends Euler A

Euler

The deterministic (non-ancestral) Euler sampler.

• Generally more stable and predictable than Euler A
• Tends to converge to a cleaner result at the same seed when steps are increased
• Works at typical step counts around 20-30

Use this when:

• You want Euler's overall feel but with more stability than Euler A
• You are reproducing settings from a workflow that uses plain Euler

LCM

Made for very low step counts and fast generation.

• Best when speed matters more than maximum detail
• Usually paired with very low steps, often around 4-8
• At high step counts it often gives little benefit
• Karras does not apply to LCM

Use this when:

• You want quick previews
• You are using an LCM-tuned workflow or model

If results look weak or strange with LCM, switch back to DPM++ 2M.

Karras

The Karras toggle changes the noise schedule rather than the sampler math. Switching it on or off keeps the same scheduler family but redistributes how noise is removed across steps.

• Off: the scheduler's default sigma schedule
• On: the Karras sigma schedule, which spends more steps refining low-noise (later) regions
• Most visible at moderate step counts (around 20-30); at very high step counts the difference shrinks
• Not available for LCM

If you are matching settings from a model page, check whether the author wrote something like DPM++ 2M Karras — if so, pick DPM++ 2M and turn Karras on.

Steps

Range: 1-50

Steps are the number of denoising iterations. More steps give the model more chances to refine the image, but more is not always better.

• Too low: blurry, incomplete, weak detail
• Good range for most use cases: 20-30
• Very high: slower, with smaller gains after a point

Beginner advice:

• Start at 24
• If the image looks unfinished, try increasing to 28 or 30
• If you are using LCM, use much lower values instead

CFG

Range: 1-30

CFG means classifier-free guidance. You can think of it as prompt strength.

• Lower CFG: the model has more freedom; images may look natural but less faithful to the prompt
• Higher CFG: the model obeys the prompt more strictly, but can look harsh, overcooked, or unnatural

For beginners, 5-8 is the safest range.

Common behavior:

• 1-4: loose prompt adherence
• 5-8: balanced and usually best
• 9-14: stronger prompt obedience, but watch for artifacts
• 15+: often excessive unless a specific model recommends it

If an image ignores an important subject from your prompt, raise CFG slightly. If colors, faces, or lighting start looking unnatural, lower it.

Prompt quality still matters more than extreme CFG. If your prompt is vague or contradictory, raising CFG usually will not fully fix it.

TIP

Performance note: On the NPU path, setting CFG to exactly 1 skips the unconditional UNet pass entirely, roughly halving the time per step. This is especially useful with LCM-distilled models that work well at CFG 1.

Denoise Strength

Range: 0-1

This setting mainly matters for img2img, not normal txt2img. It controls how much the output is allowed to change away from the input image.

• 0: almost keeps the original image unchanged
• 1: almost rebuilds the image from scratch

Practical ranges:

• 0.2-0.4: small fixes, light restyling
• 0.5-0.7: noticeable edits while keeping the original structure
• 0.75-0.85: strong redraw with composition mostly preserved
• 0.9-1.0: very large changes, composition may drift

For Local Dream's SD1.5 NPU high-resolution workflow, around 0.8 is a strong default. See SD1.5 NPU High Resolution.

Seed

Range: uint32 (0 to 4294967295)

The seed is the random starting point for generation. Using the same model, prompt, and settings with the same seed helps you reproduce a result.

• Change the seed when you want new compositions
• Keep the seed when you want to compare prompt or parameter changes fairly

Example:

• You like the pose but want a different outfit: keep the same seed first
• You want totally different image ideas: change the seed

For more detail about reproducibility across CPU, GPU, and NPU, see Seed Settings.

How To Tune Without Getting Lost

For beginners, this order works well:

1. Pick a model and check the author's example settings if available.
2. Write a simple prompt, and add a small negative prompt only if needed.
3. Keep the scheduler at DPM++ 2M.
4. Set steps to around 24.
5. Set CFG to around 6 or 7.
6. Only after that, experiment with seed changes or img2img denoise strength.

If something looks wrong, try this quick checklist:

• The subject is unclear: simplify the prompt and make the main subject earlier
• There are specific defects: add a short negative prompt for those defects
• Too blurry: raise steps a little
• Not following the prompt: raise CFG a little
• Looks oversharpened or strange: lower CFG
• img2img changed too much: lower denoise strength
• img2img changed too little: raise denoise strength

If the result is still off even after small tweaks, do not keep pushing random settings higher and higher. Go back to the model's official release page and compare your prompt, negative prompt, scheduler, steps, CFG, clip skip, and prompt style against the author's example.