How SDXL prompts work
- SDXL handles longer, more descriptive prompts and natural phrasing better than SD 1.5 — lead with the subject, then style and medium, then fine detail.
- Quality boosters still help: a short "masterpiece, best quality, highly detailed" up front, plus a solid negative prompt (worst quality, low quality, blurry).
- Attention weighting is identical to SD 1.5 — (golden hour:1.3) — but keep values moderate (0.8–1.3); SDXL distorts faster at extremes.
- Many SDXL checkpoints (Juggernaut, RealVis, DreamShaper) have their own preferred tokens — save them once as snippets and stop retyping.
Example SDXL prompt
masterpiece, best quality, a lone lighthouse on a rocky cliff at golden hour, dramatic storm clouds, (volumetric light:1.2), ultra-detailed, 35mm photograph
negative: worst quality, low quality, blurry, oversaturated, deformed, extra fingers
Build SDXL prompts faster with Prompt Builder
- One-click quality and negative blocks tuned for SDXL — stamp them in, don’t retype.
- Wrap any token in attention weights with Ctrl/Cmd + ↑/↓ — no manual parentheses.
- Drag in any SDXL PNG from Automatic1111 or ComfyUI to recover its exact prompt.
- Model and LoRA filters show only the snippets relevant to your current setup.
Frequently asked questions
What’s the best prompt structure for SDXL?▾
Subject first, then style and medium, then fine detail and a short quality tag. SDXL rewards natural, descriptive phrasing more than keyword spam.
Do I still need quality tags in SDXL?▾
They help but matter less than in SD 1.5. A short "masterpiece, best quality" plus a clean negative prompt is usually enough.
Does SDXL use the same weighting syntax?▾
Yes — (token:1.3) works. Keep values moderate (roughly 0.8–1.3); SDXL distorts faster than SD 1.5 at extreme weights.