Stable Diffusion
Mastery.
Token weighting, negative prompting, and photorealism. The definitive guide to creating production-ready imagery with Stable Diffusion SDXL and beyond.
Affiliate link, we may earn a commission at no extra cost to you.
The Physics of an Elite Image Prompt
Stable Diffusion doesn't see sentences, it sees tokens. To rank as a power user, your prompt structure must follow hierarchical logic: [Subject] + [Action/Pose] + [Environment/Background] + [Lighting/Mood] + [Camera/Technical] + [Artist Style].
The Prompt Hierarchy
- 1. Subject & Action (specific materials, weathering)
- 2. Technical Camera Specs (85mm f/1.8, 35mm f/8, 14mm)
- 3. Lighting & Atmosphere (volumetric fog, ray tracing, rim lighting)
- 4. Advanced Syntax (token weighting, negative prompts, wildcards)
- 5. Artist Style (photographic, painterly, cinematic)
Part 1 of 2: Stable Diffusion Mastery
This guide covers the fundamentals: 80+ photorealistic prompts, token weighting, and the negative prompt cheatsheet. Ready for ControlNet, LoRA stacking, and production SDXL workflows? Continue to the Stable Diffusion Advanced Guide once you have the basics locked in.
Token Weighting: The Power Feature
Token weighting uses (keyword:1.3) to increase emphasis or (keyword:0.7) to reduce it. For example, (cyberpunk:1.4) ensures the neon aesthetic dominates.
A rugged fisherman with (weathered skin:1.4) and (salt-and-pepper beard:1.2), (oil painting:0.8), professional portrait
Technical Camera Specifications
To achieve photorealism, you must prompt for the lens. Each focal length produces dramatically different results.
85mm f/1.8
Perfect for portraits with creamy background blur (bokeh)
35mm f/8
Ideal for sharp street photography and environmental shots
14mm Wide-Angle
Used for grand architecture or sprawling landscapes
Lighting & Atmosphere: The Soul of the Image
Lighting dictates the "soul" of the image. Use technical terms like "Volumetric Fog," "Ray Tracing," "Global Illumination," or "Rim Lighting" to add 3D depth.
Advanced Syntax & Power Features
Negative Prompting
Remove the "AI look" with essential negative tokens: blurry, low contrast, distorted limbs, extra fingers, cartoonish, oversaturated.
Wildcards & Randomized Choice
Use brackets a warrior | a mage | a thief to generate variations in a single batch, vital for character design.
Ready to go deeper? Our Stable Diffusion Advanced guide covers ControlNet, LoRA stacking, regional prompting and production-grade SDXL workflows that build on everything in this page.
10 SDXL Prompt Recipes That Actually Work in 2026
These ten recipes are the workhorse formulas SDXL artists run in production in April 2026. Each one pairs a positive prompt structure with the negative-prompt seasoning that prevents the failure mode the recipe is most prone to. Copy, paste, swap the bracketed variables, and run at 1024x1024 with DPM++ 2M Karras at 28 steps unless noted.
1. Editorial portrait
cinematic editorial portrait of a [SUBJECT], (sharp eyes:1.2), (skin pores:1.1), 85mm f/1.4, golden hour, rim lighting, kodak portra 400, photographed by Annie Leibovitz
Negative: blurry, plastic skin, smooth, airbrushed, oversaturated, amateur, distorted hands. Pair with the FaceDetailer extension for clean eyes.
2. Product hero shot
studio product photography of [PRODUCT], (clean white background:1.3), softbox lighting, 100mm macro, sharp focus, commercial e-commerce, hyperreal materials, no shadows
Negative: cluttered, props, hand, person, model, dramatic lighting, vignette. Set CFG to 4-5 to keep the look clean.
3. Cinematic environment
cinematic still of [LOCATION], (volumetric fog:1.3), god rays, anamorphic lens, 35mm film grain, color graded teal and orange, shot on Arri Alexa, blade runner 2049 reference
Negative: cartoon, illustration, low contrast, oversharp, fisheye. Render at 1216x832 for the 3:2 cinematic frame.
4. Concept art / matte painting
concept art of [SCENE], (epic scale:1.3), painterly brushstrokes, atmospheric perspective, by Greg Rutkowski and Craig Mullins, ArtStation trending, 8k
Negative: photographic, photorealistic, bokeh, lens flare, signature, watermark. Add (artistic:1.2) to push it further from photorealism.
5. Anime / illustration
anime key visual of [CHARACTER], (clean lineart:1.2), cel shading, vibrant flat colors, by Kyoto Animation, official artwork, pixiv ranking
Negative: 3d render, photograph, realistic, blurry lines, sketch, jpeg artifacts. Use the Anything-V5 or Counterfeit checkpoint for stronger anime fidelity.
6. Architectural exterior
architectural photography of a [BUILDING TYPE], (golden hour light:1.2), 14mm wide angle, hdr, sharp focus, professional real estate, dezeen feature
Negative: people, cars, distorted perspective, fisheye, motion blur. Use ControlNet MLSD for clean straight lines on architectural subjects.
7. Food photography
overhead food photography of [DISH], (steam rising:1.2), shallow depth of field, natural window light, rustic wooden table, bon appetit magazine, 50mm f/2.8
Negative: plastic, fake, oversaturated, distorted, low quality, watermark. Avoid (mouth:1.0), SDXL renders human anatomy adjacent to food awkwardly.
8. Sci-fi character
full body shot of a [CHARACTER] in [SCI-FI ARMOR], (intricate mechanical details:1.3), volumetric backlight, neon accent lighting, cinematic, by Maciej Kuciara
Negative: deformed, extra arms, asymmetrical armor, low detail, blurry, cartoon. Best paired with a sci-fi LoRA at 0.6-0.8 strength.
9. Game asset / icon
game icon of a [ITEM], (centered composition:1.3), clean black background, soft glow, polished metal materials, high contrast, mobile game ui asset
Negative: scene, environment, person, hands, shadow on background, watermark. Render square at 1024x1024 for direct sprite use.
10. Brand mockup
flat lay brand identity mockup with [PRODUCT], minimal stationery, monochrome palette, soft shadow, top-down 90 degree, behance feature, agency portfolio
Negative: cluttered, neon, busy, gradient, low quality. Excellent for landing-page hero imagery for SaaS launches.
The Negative Prompt Cheatsheet by Use Case
A great positive prompt and a sloppy negative prompt is still a sloppy image. The negative prompt is where 80% of the “does this look like AI slop?” problem gets fixed. Below are the eight tuned negative-prompt strings used across the SDXL community on Civitai and ComfyUI workflows in April 2026.
Photorealism (universal)
blurry, low quality, watermark, logo, text, signature, distorted, bad anatomy, extra limbs, extra fingers, missing fingers, oversaturated, plastic, amateur, cartoon, illustration, 3d render, painting, drawing
Portrait fidelity
deformed face, asymmetrical eyes, lazy eye, dead eyes, mouth open weirdly, missing teeth, plastic skin, airbrushed, smooth skin, doll-like, mannequin, uncanny valley, off-center, bad proportions
Product / e-commerce
person, hand, finger, cluttered background, props, dramatic lighting, vignette, watermark, logo, text, low quality, blurry, motion blur, jpeg artifacts
Architectural
people, cars, motorcycles, distorted perspective, fisheye, motion blur, low resolution, double exposure, watermark, oversaturated, hdr halo, lens flare
Anime / illustration
3d render, photograph, photorealistic, blurry lines, sketch, jpeg artifacts, lowres, bad anatomy, bad hands, signature, watermark, ugly, deformed
Concept art
photograph, photorealistic, bokeh, lens flare, signature, watermark, low contrast, smudged, sloppy, low effort, deviantart 2010, low resolution
Game asset / icon
scene, environment, person, hands, shadow on background, watermark, signature, blurry, low quality, off-center, cropped
Cinematic still
cartoon, illustration, low contrast, oversharp, fisheye, distorted, blurry, lowres, signature, watermark, painting, anime
Negative prompts work best when paired with a moderate CFG scale of 5-7. Going above CFG 9 amplifies both your positive and negative prompts, which can over-correct and produce visible artifacts (waxy skin, color banding, doubled eyes). If your negative prompt feels overpowering, drop CFG to 4.5 before adding more tokens. For a deeper dive into ControlNet, IP-Adapter and LoRA stacking, the Stable Diffusion Advanced guide extends every recipe on this page.
Sampler & Step Count Decision Matrix
Choosing the right sampler is the difference between a 9-second generation and a 27-second generation, and it changes the visual character of the image. Here is how SDXL artists pick samplers in 2026.
| Sampler | Steps | Best for | Trade-off |
|---|---|---|---|
| DPM++ 2M Karras | 28-32 | Default workhorse, photorealism, portraits | Slightly slower than Euler |
| Euler a | 20-25 | Speed runs, quick prototypes, anime | Slightly noisier in small details |
| DPM++ SDE Karras | 25-30 | Maximum detail, hero shots, hi-res fix | ~1.4x slower than DPM++ 2M |
| UniPC | 15-20 | Fastest path to acceptable quality | Less consistent across seeds |
| DPM++ 3M SDE Karras | 30-40 | Studio-quality finals before upscale | Slowest, but visibly cleaner |
| LCM (Latent Consistency) | 4-8 | Real-time preview, batch ideation | Lower fidelity than full samplers |
The default that ships clean results 90% of the time on SDXL is DPM++ 2M Karras at 28 steps with CFG 6. Use UniPC at 15 steps for prompt iteration to keep generation under 6 seconds, then re-render the winner at DPM++ SDE Karras 28 steps for the final asset.
The Photorealism Checklist 2026
To ensure your images aren't flagged as "synthetic," include these high-frequency keywords in your workflow:
- ✓ Subsurface scattering
- ✓ 4k raw photo
- ✓ Unfiltered
- ✓ Micro-texture
- ✓ High dynamic range (HDR)
- ✓ Ray-traced shadows
Ready to go deeper?
Advanced Stable Diffusion: ControlNet, LoRA Stacking & Pro Negative Prompt Recipes
You've mastered the basics. Our advanced guide covers ControlNet pose transfer, LoRA stacking for custom styles, and the negative prompt recipes that power users use to eliminate every artifact, distorted hands, watermarks, blurry backgrounds, and more.
Read the Advanced Guide →