SDXL Prompting Guide

A practical guide for realism, photorealism, and anime (including adult-only content). Learn how positives/negatives, weights, CFG/steps, samplers, and schedulers interact — and how to fix “bad results” fast.
Realism / Photoreal
Anime / Illustration
Troubleshooting
Safety rules
Safety: Attempts to generate illegal content (especially anything involving minors) are prohibited and may be logged and escalated where legally required. If you’re here for that — stop. This guide only covers legitimate adult / artistic use.
Tip: Use the search to filter sections instantly.
Support Submit Ticket
Find topics

1) Prompt structure that actually works

The easiest way to get consistent results is to write prompts in a repeatable “stack”. Think: Subject → Attributes → Environment → Lighting → Camera → Style/Quality → Constraints.
Recommended “stack” ▾
  • Subject: who/what is it?
  • Attributes: age (adult), hair, clothing, mood, action, details that matter
  • Environment: location + era + props
  • Lighting: softbox, window light, neon rim, golden hour, etc.
  • Camera: lens, angle, depth of field, focus
  • Style: photorealistic / cinematic / anime key visual / painterly
  • Constraints: “single subject”, “looking at camera”, “hands visible”, “no text”
Common mistake: adding 50 style buzzwords that fight each other. If your prompt says “photorealistic” AND “anime” AND “oil painting”, the model picks one and gets confused.
Template you can reuse 
[SUBJECT], [AGE=adult], [KEY FEATURES], [ACTION/POSE],
[LOCATION/ENV], [TIME OF DAY], [LIGHTING],
[camera: lens, angle, depth of field],
[STYLE], [QUALITY], [CONSTRAINTS]

2) Weights & emphasis (control the model)

Use weights to force what matters and weaken what doesn’t. Don’t overdo it — small changes are huge.
How to emphasize or reduce concepts ▾
  • (keyword) = mild emphasis
  • (keyword:1.2) = strong emphasis
  • (keyword:1.4) = very strong (can cause artifacts)
  • (keyword:0.8) = de-emphasize
Rule of thumb: if you need 1.6+ weights everywhere, the prompt is fighting itself. Simplify the prompt first.
What to weight (and what NOT to) ▾
  • Good to weight: “single subject”, “looking at camera”, “hands visible”, “realistic skin texture”, “sharp focus”
  • Risky to weight heavily: “extremely detailed”, “8k”, “masterpiece” (often increases fake-looking artifacts)
  • Don’t weight everything: weights are for priorities, not decoration
Example weights 
(single subject:1.2), (looking at camera:1.15), (sharp focus:1.1),
(realistic skin texture:1.15), (hands visible:1.1)

3) Negatives that actually help

Negatives are for removing failure modes. Too many negatives can make images “dead” or plastic. Use a core negative and add targeted negatives only when needed.
Core negative (general SDXL quality) ▾
Core negative 
bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, deformed hands,
cross-eye, lazy eye, deformed face, asymmetrical eyes, blurry, lowres, noise, jpeg artifacts,
text, watermark, logo, signature, cropped, out of frame
Targeted negatives (only when needed) ▾
  • Over-smoothed skin / “plastic”: add overprocessed, airbrushed, waxy skin
  • Weird eyes: add pupils mismatch, iris deformation
  • Bad teeth/mouth: add bad teeth, deformed mouth
  • Background clutter: add busy background, clutter
  • Over-sharpening: add oversharpened, harsh edges
If your image gets dull after adding negatives, remove half of them. SDXL often prefers clear positives over massive negatives.

4) CFG & Steps: the sweet spots

CFG controls how hard the model follows your text. Steps control how much sampling time you give it. Most “crappy” results are just wrong CFG/steps for the sampler/model.
Practical ranges (fast + quality) ▾
  • CFG: usually 4.5 – 7.5 for SDXL. Higher can cause “burned” / crunchy artifacts.
  • Steps: usually 20 – 35 (depends on sampler). Too high can overcook textures.
  • Common mistake: CFG 12+ with SDXL — often ruins realism and skin.
Fast default: CFG ~6, steps ~28, DPM++ 2M / SDE + Karras (if available).
When to change steps vs CFG ▾
  • If it ignores prompt details: increase CFG slightly (e.g., +0.5), or improve the prompt stack.
  • If it’s noisy / unfinished: increase steps by +6 to +10.
  • If it’s “crunchy / overbaked”: reduce CFG by -0.5 or steps by -6.

5) Samplers: what they’re good at

Samplers strongly affect “look” and consistency. Here are the common ones in SDXL/Forge UI and what to expect.
Sampler Good at Bad at / risks Suggested use
DPM++ 2M Clean detail, stable structure, great all-round Can look slightly “too perfect” at high CFG Default for realism + general prompts
DPM++ SDE Natural textures, strong realism, good depth Needs decent steps; too low steps = mush Photoreal / cinematic scenes
DPM++ 3M Very detailed, strong adherence Can overcook; more artifacts if pushed When you need extra detail
Euler a Stylized punch, anime/illustration vibes Can get noisy/chaotic for photoreal Anime / stylized / “edgy” looks
Euler Stable, simpler look, good baseline Less rich detail than DPM++ in many cases Fast drafts / simple compositions
DDIM Predictable, gentle, consistent Can look flat, less micro-detail When you want calm/consistent outputs
Heun Smooth gradients, stable renders Slower feel; sometimes less “punch” Portraits with smooth lighting

Sampler + Scheduler Wizard (auto recommendations)

Pick your goal and priority — this suggests a solid sampler, scheduler, steps, and CFG. Use it as a starting point, then tweak one knob at a time.
Reminder: If your result looks “anime-ish” when you want photoreal, switch away from Euler a, lower CFG, and remove stylized keywords.
Goal
Priority
Subject type
Look
Recommended settings
Click “Recommend” to generate suggestions…
How to use: Try the recommended combo for 3–5 gens. If it’s crunchy → lower CFG. If it’s unfinished → add steps. If it ignores details → small CFG increase (+0.5) or simplify your prompt.
If your image looks “anime-ish” when you want photoreal, try switching from Euler a to DPM++ 2M / SDE, reduce CFG, and remove stylized keywords.

Prompt Builder (positive + negative + LoRAs)

Build a clean prompt stack without forgetting the important pieces. Pick a preset, fill a few fields, and copy the final output.
Note: Keep prompts consistent. If you pick Photoreal, avoid adding “anime/illustration/cel shading”.
Preset
Quality strength
Subject
Wardrobe / accessories
Location / environment
Lighting / camera
Extra details (optional)
Tip: use weights like (freckles:0.6) or (looking at camera:1.15).
LoRA helpers (optional)
LoRA intensity
Render settings preset
Positive prompt
Choose options then click “Build prompt”…
Negative prompt
—
Settings line
—
Reminder: Do not attempt illegal content. Do not include minors or “youth-coded” terms. These attempts may be logged and escalated.

6) Schedulers: stability vs punch

Schedulers shape the noise curve over steps. If you have them, use them — they matter.
Scheduler Feel Best for Notes
Karras Balanced + stable Most SDXL work Usually the safest default with DPM++
Exponential More contrast/punch Stylized, dramatic lighting Can exaggerate textures
SGM Uniform Even progression Consistency / smoothness Often good for portraits & calm looks
Simple / Normal Plain baseline Compatibility If your model behaves weird on other schedules
If you don’t know what to pick: Karras + DPM++ 2M is a strong SDXL default.

7) Photoreal “recipe” (strong default)

Photoreal comes from: clean prompt stack + realistic lighting/camera + avoiding stylized buzzwords + sane CFG.
Photoreal prompt (copy/paste) ▾
Positive (Photoreal) 
adult woman, single subject, natural expression, looking at camera,
realistic skin texture, subtle pores, fine hair strands, natural makeup,
soft window light, cinematic soft shadows, shallow depth of field,
85mm lens, f/1.8, sharp focus on eyes, bokeh background,
photorealistic, high detail, clean color grading, realistic
Negative (Photoreal) 
bad anatomy, bad hands, extra fingers, missing fingers, fused fingers, deformed hands,
cross-eye, asymmetrical eyes, deformed face, blurry, lowres, noise, jpeg artifacts,
text, watermark, logo, signature, overprocessed, airbrushed, waxy skin
Suggested settings: DPM++ 2M or DPM++ SDE • Karras • steps 26–34 • CFG 5.5–7

8) Anime / stylized “recipe”

Anime wants clarity: character + outfit + scene + lighting + style keywords that match the model.
Anime prompt (copy/paste) ▾
Positive (Anime) 
anime illustration, single character, adult woman,
clean lineart, expressive eyes, detailed hair, soft cel shading,
vibrant color palette, dramatic rim light, soft background,
dynamic composition, studio anime key visual, high quality
Negative (Anime) 
realistic photo, uncanny realism, bad anatomy, bad hands, extra fingers, missing fingers,
lowres, blurry, jpeg artifacts, text, watermark, logo, signature
Suggested settings: Euler a (or DPM++ 2M) • steps 22–32 • CFG 6–8

9) Adult-only prompts (18+)

Keep adult prompts clean: describe lighting, pose, camera, and quality. Avoid spammy keyword piles and never include anything illegal.
How to keep adult results high quality ▾
  • Always specify adult explicitly if your platform requires it.
  • Use camera + lighting to control vibe more than raw explicit words.
  • If things go weird: lower CFG slightly and simplify the prompt.
Never attempt anything involving minors or “youth-coded” language. Accounts may be banned and actions logged.

10) Composition & camera control

Composition words are “cheat codes” for predictable framing.
Useful camera phrases ▾
  • Framing: close-up, headshot, bust shot, half-body, full body, wide shot
  • Angle: eye-level, low angle, high angle, over-the-shoulder
  • Lenses: 35mm (environment), 50mm (natural), 85mm (portrait), 135mm (compressed/bokeh)
  • DOF: shallow depth of field, bokeh, subject in focus, background blurred
  • Lighting: soft window light, rim light, neon, golden hour, diffused softbox
Camera add-on pack 
85mm lens, shallow depth of field, sharp focus on eyes, bokeh background,
soft window light, cinematic soft shadows, natural color grading

11) Hands & faces: practical fixes

Hands are hard. The goal is to reduce chances of failure and keep them readable.
Hands: improve success rate ▾
  • Add: hands visible, fingers, natural hand pose
  • Don’t force impossible poses (twisted wrists, extreme foreshortening)
  • Full body + hands far away = tiny hands = more mistakes
  • If your UI supports it: do an inpaint pass on hands only
Eyes/faces: keep them consistent ▾
  • Add: sharp focus on eyes, symmetrical face
  • Lower CFG slightly if you get “warped eyes”
  • Use less “ultra-detailed” buzzwords

12) “My image came out crappy” (diagnose fast)

Most failures are predictable. Here’s the fix list.
Blurry / muddy ▾
  • Increase steps +6 to +10
  • Switch to DPM++ 2M / SDE
  • Add sharp focus, specify lens/DOF
Too “anime” when you wanted photoreal ▾
  • Remove “anime/illustration/masterpiece” style words
  • Switch sampler: Euler a → DPM++ 2M/SDE
  • Lower CFG by ~0.5 to 1.5
  • Add camera + lighting realism words (lens, bokeh, skin texture)
Crunchy / over-sharpened / “burned” ▾
  • Lower CFG
  • Reduce steps slightly
  • Remove “8k ultra detailed HDR masterpiece” spam
  • Add negative: oversharpened, harsh edges
Wrong pose or wrong clothing ▾
  • Simplify prompt and weight the important part: (red dress:1.2)
  • Use fewer conflicting clothing descriptors
  • Try a different seed or slightly different sampler

13) Advanced knobs (when you want pro-level control)

These are powerful, but only worth it once your base prompting is solid.
Prompt order & token budget ▾
  • Put the most important concept early in the prompt.
  • If your prompt is huge, the model may ignore late details.
  • Cut anything that isn’t doing visible work.
Hi-res / 2-pass rendering (if your UI supports it) ▾
  • Base pass builds structure; hi-res pass adds detail.
  • Denoise too high can “change the person.” Start modest.
  • For realism, less is more: keep it natural.
Refiner (SDXL base → refiner) ▾
  • Refiner can improve micro-detail, but can also over-smooth.
  • If you’re getting plastic skin, reduce refiner strength or skip it.

Quick Presets (model-based defaults)

Pick your model type and goal — this gives you a strong starting point for settings. If your output looks off, change one knob at a time.
Model type
Priority
Target
Look
Optional photoreal LoRA pack
These are best on photoreal models. If you don’t know, leave on Sweet spot.
LoRA intensity
Resolution bias
Preset output
Choose options then click “Generate preset”…
Want the fastest upgrade in quality? Use a cleaner prompt, pick the right sampler, and stop overusing buzzwords. Most “bad” images are prompt conflicts, not “the AI”.
Terms & Policies: Terms of Service • Privacy Policy
Support Still need help? Submit a ticket