What settings should I use for SDXL?

Start from what our corpus converges on: 30 steps, CFG 6.5, and the DPM++ 2M Karras sampler. Fix your seed the moment a composition lands so later tweaks stay comparable. CFG around 6–7 keeps the model faithful without over-baking contrast; higher values tend to harden and posterize. Adjust from this baseline one variable at a time rather than changing several at once.

Why does SDXL need a negative prompt?

SDXL treats the negative prompt as a first-class channel — 100% of our tested SDXL prompts rely on a universal baseline that guards common failure modes like mangled text, extra fingers and artifacts. It is the single thing no other major model needs. Extend the baseline with whatever you actually see going wrong, named literally: doubled wings, extra heads, watermark ghosting.

How do I fix recurring artifacts in SDXL images?

See the artifact, name the artifact. When a specific flaw keeps appearing — extra heads, doubled wings, watermark ghosting — add that exact term to your negative prompt rather than rewording the positive prompt. Keep steps at 30 and CFG near 6.5, fix the seed, and change one thing per run so you can tell which edit removed the problem.

Stable Diffusion Prompt Builder

SDXL responds to comma-separated visual layers plus one thing no other major model needs: a negative prompt. This builder assembles the positive side — subject, style, lighting, framing, palette — and pairs it with the universal negative baseline that 100% of our tested SDXL prompts rely on. Copy both, load your checkpoint, and generate.

Subject (describe what you want) Style Lighting Framing Palette Aspect ratio

Your prompt

Fill in a subject above — the prompt assembles here as you choose.

Negative prompt (baseline — edit freely)

text, watermark, logo, blurry, lowres, jpeg artifacts, deformed, extra fingers, extra limbs, bad anatomy

Suggested settings: 30 steps · CFG 6.5 · DPM++ 2M Karras

How to use it well

Start from the settings our corpus converges on — 30 steps, CFG 6.5, DPM++ 2M Karras — and fix your seed the moment a composition lands, so later tweaks stay comparable. The negative baseline above guards the universal failure modes (mangled text, extra fingers, artifacts); extend it with what you actually see going wrong, named literally: doubled wings, extra heads, watermark ghosting. That diagnostic habit — see the artifact, name the artifact — is the core skill, and our negative prompts guide teaches it with copy-ready lists per use case.

Every style in the dropdown is backed by tested entries with published settings — the anime and pixel art hubs are where SDXL particularly shines, and the full SDXL library shows the negative prompt each image actually required. Builders draft; tested prompts deliver.

Frequently asked questions

What settings should I use for SDXL?: Start from what our corpus converges on: 30 steps, CFG 6.5, and the DPM++ 2M Karras sampler. Fix your seed the moment a composition lands so later tweaks stay comparable. CFG around 6–7 keeps the model faithful without over-baking contrast; higher values tend to harden and posterize. Adjust from this baseline one variable at a time rather than changing several at once.
Why does SDXL need a negative prompt?: SDXL treats the negative prompt as a first-class channel — 100% of our tested SDXL prompts rely on a universal baseline that guards common failure modes like mangled text, extra fingers and artifacts. It is the single thing no other major model needs. Extend the baseline with whatever you actually see going wrong, named literally: doubled wings, extra heads, watermark ghosting.
How do I fix recurring artifacts in SDXL images?: See the artifact, name the artifact. When a specific flaw keeps appearing — extra heads, doubled wings, watermark ghosting — add that exact term to your negative prompt rather than rewording the positive prompt. Keep steps at 30 and CFG near 6.5, fix the seed, and change one thing per run so you can tell which edit removed the problem.