Stable Diffusion
Prompt Generator.
Free builder for SDXL, SD 1.5, and SD 3. Outputs positive + negative + sampler + CFG + steps in one paste-ready block.
Works with Automatic1111, ComfyUI, Invoke AI, and most community fine-tunes.
Describe what you want
3 prompt variations
Click Copy to use# POSITIVE PROMPT [describe your subject], photorealistic # NEGATIVE PROMPT lowres, bad anatomy, bad hands, text, watermark, signature, blurry, jpeg artifacts, out of frame, duplicate, extra fingers, mutated hands, deformed face, asymmetric eyes # SETTINGS Sampler: DPM++ 2M Karras CFG scale: 7 Steps: 30 Note: Good for iteration — raises CFG to 9 if subject wanders.
# POSITIVE PROMPT [describe your subject], photorealistic, masterpiece, best quality, ultra-detailed, 8k, award-winning photography, sharp focus, intricate detail # NEGATIVE PROMPT lowres, bad anatomy, bad hands, text, watermark, signature, blurry, jpeg artifacts, out of frame, duplicate, extra fingers, mutated hands, deformed face, asymmetric eyes # SETTINGS Sampler: DPM++ 2M Karras CFG scale: 7 Steps: 30
# POSITIVE PROMPT [describe your subject], photorealistic, masterpiece, best quality, ultra-detailed, 8k, award-winning photography, sharp focus, intricate detail, trending on artstation, highly detailed, volumetric lighting, film grain, cinematic color grading, anamorphic lens, depth of field, volumetric lighting, detailed background # NEGATIVE PROMPT lowres, bad anatomy, bad hands, text, watermark, signature, blurry, jpeg artifacts, out of frame, duplicate, extra fingers, mutated hands, deformed face, asymmetric eyes # SETTINGS Sampler: DPM++ 2M Karras CFG scale: 7 Steps: 30 Note: Drop steps to 25 and sampler to Euler a for faster iteration.
Under the hood
Why Stable Diffusion rewards structured prompts.
Front-load the subject. SD weights early tokens heavier than later ones, so a strong first clause ('portrait of a weathered sea captain') anchors the model before style tokens dilute its attention.
The negative prompt is where you fight common SD failure modes — bad hands, extra fingers, blurry, jpeg artifacts. The generator tailors these to your image type so you don't have to remember them all.
CFG controls how literally the model follows your prompt. Sampler controls how noise gets denoised. Steps control how many refinement passes. Together these produce the difference between a usable image and a hallucinated mess.
Related free tools
Specialized generators for specific tasks.
Midjourney Prompt Generator
v6 / niji 6 syntax with --ar, --stylize, --style raw flags.
Flux Prompt Generator
Black Forest Labs Flux — high prompt adherence.
AI Art Prompt Generator
Model-agnostic art prompts across medium and style.
Stable Diffusion Guide
Full guide to models, LoRAs, ControlNet, and workflows.
FAQ
Questions about Stable Diffusion prompting.
Does this work for SDXL, SD 1.5, and SD 3?+
Yes. The prompt structure is version-agnostic — subject, style, artist influence, and quality boosters work across Stable Diffusion 1.5, SDXL, SD 3, and most community fine-tunes. Adjust CFG slightly lower on SDXL (5-7) vs SD 1.5 (7-9).
What's the auto-generated negative prompt doing?+
The generator adapts the negative prompt to your image type. Portraits get anatomy terms (bad hands, extra fingers, deformed face). Landscapes exclude figures. Product shots exclude logos and busy backgrounds. You can append your own terms in the extra negative field.
What CFG scale should I use?+
Start at 7. If the output ignores your prompt, raise to 9-11. If it looks over-baked, burned, or contrasty, drop to 5-6. SDXL tolerates lower CFG (4-6) better than SD 1.5 because it was trained on a refiner pass.
Which sampler is best?+
DPM++ 2M Karras is the safe default — clean, fast, works at 25-30 steps. Euler a is faster but noisier at low steps. UniPC converges in 15-20 steps. DDIM is useful when you need deterministic seed behavior for animations.
Why does the generator include artist names?+
Artist style tokens (Greg Rutkowski, Artgerm, Moebius) produce strong stylistic anchors in SD 1.5 and SDXL. For SD 3, artist tokens work less reliably — prefer descriptive style language ('oil painting with visible brushstrokes, dramatic chiaroscuro') instead.
Can I paste this into Automatic1111, ComfyUI, or Invoke AI?+
Yes. The positive prompt goes in the main prompt field. The negative prompt goes in the negative field. Sampler, CFG, and steps paste into the standard settings slots. In ComfyUI, the positive/negative feed into the CLIP text-encode nodes.
What about weighted prompts and token attention?+
The generator outputs clean prompts without weights by default. To emphasize a term, wrap it in parentheses: (volumetric lighting:1.3). To de-emphasize, use square brackets or weights below 1.0: (background:0.6). Only add weights when a specific element is being ignored or overpowered.
Is the prompt portable to Midjourney or Flux?+
The descriptive body transfers well. Strip the SD-specific negative prompt and settings block — Midjourney and Flux don't use them. For Midjourney, use our dedicated generator for proper --ar / --v syntax.