Stable Diffusion prompt + negative prompt masterclass

PROMPT: [SUBJECT], [DETAILS], [STYLE], [LIGHTING], masterpiece, best quality, highly detailed, 8k

NEGATIVE PROMPT: lowres, bad anatomy, extra fingers, missing fingers, fused fingers, watermark, text, signature, blurry, deformed, jpeg artifacts, oversaturated, mutilated, poorly drawn face, extra limbs

What each negative token actually prevents:
• 'bad anatomy, extra fingers, fused fingers' — targets the most common SDXL failure: hands. Always include these.
• 'watermark, text, signature' — training data is full of watermarked images; without these, logos bleed through.
• 'blurry, lowres, jpeg artifacts' — forces the model toward the sharp, high-quality end of its training distribution.
• 'oversaturated' — prevents the garish neon colors that models default to when you ask for 'vibrant.'
• 'poorly drawn face, deformed' — general anatomical safety net.

Settings explained:
• Steps 28-35: Below 25, detail is undercooked. Above 40, you get diminishing returns and risk over-frying textures. Start at 30.
• CFG 6-8: This is prompt adherence strength. CFG 5 = loose/creative (good for art). CFG 7 = balanced. CFG 10+ = literal but can look artificial and over-sharpened. If your output ignores your prompt, raise CFG. If it looks crunchy or over-processed, lower it.
• DPM++ 2M Karras: This sampler converges faster than Euler — meaning it produces clean results in fewer steps. Karras noise schedule front-loads detail and smooths out at the end. For photorealism, also try DPM++ SDE Karras.
• Hires fix: Essential for faces. Generate at 512x512 (SD1.5) or 1024x1024 (SDXL), then upscale with hires fix to avoid the 'melted face at high res' problem.

Example 1 — Photorealism: 'Portrait of a weathered fisherman, deep wrinkles, salt-and-pepper beard, golden hour coastal light, shot on 85mm lens, shallow depth of field, masterpiece, best quality, 8k'

Example 2 — Anime style (different settings): 'A mage casting a spell in a crystal cavern, dynamic pose, glowing particles, fantasy anime style, dramatic purple and teal lighting, masterpiece, best quality'
For anime: lower CFG to 5-6, use Euler a sampler, and ADD to the negative prompt: 'photorealistic, 3d render, photograph' to prevent style bleed.

Flux note: Flux ignores negative prompts entirely — instead, phrase what you DON'T want as 'without [X]' directly in the main prompt.