The reliable NSFW prompt formula is: subject and adult tag, then action or pose, then outfit, then setting, then lighting, then camera, then art style, then quality or score tags, and finally a strong negative line. Order matters because models weight earlier tokens more heavily. Keep baseline safety negatives on every render.
Most people who struggle with adult AI images do not have a model problem. They have a structure problem. They throw twenty ideas into a box in random order, the model weights the wrong tokens, and the result looks generic, deformed, or off concept. A prompt is not a wish list. It is an ordered instruction set, and the order is part of the instruction.
This pillar teaches the structure that works across Stable Diffusion 1.5, SDXL, Pony Diffusion, Illustrious, and Flux. Every subject described here is an adult (18+), fictional, AI-generated character. You should never prompt for a real identifiable person’s likeness, and you should never prompt for minors or minor-appearing subjects. Baseline safety negatives stay on for every single render, no exceptions.
Want to test these structures live without installing anything? Use our free generator and paste the worked examples below.
Why prompt order changes the output
Diffusion models read your prompt as a weighted sequence. Tokens near the front carry more influence over the composition than tokens buried at the end. That is why “photorealistic, 1woman” and “1woman, photorealistic” can produce noticeably different images. The first puts style in the driver seat. The second puts the subject in charge and lets style refine it.
A good formula front loads the things you care about most (who and what they are doing) and back loads the polish (quality tags, score tags). It groups related ideas so the model does not have to reconcile conflicting signals scattered across the string.
Here is the canonical order this whole cluster is built around:
- Subject plus adult tag (1woman, 1man, adult, the character identity)
- Action or pose (standing, sitting, looking at viewer)
- Outfit and clothing state
- Setting and background
- Lighting
- Camera, shot type, lens
- Art style or medium
- Quality or score tags
- Negative prompt (always last, separate field)
You do not need every slot on every image. But when something looks wrong, walking the slots in order is the fastest way to find which one the model ignored.
The nine slots explained
Slot 1: subject and adult tag
Start with the count and the subject. In booru-trained models that means 1girl or 1woman, 1boy or 1man, or 2girls. Always pair it with an explicit adult marker like adult, mature female, 25 years old. This is both a safety and a quality decision. Models drift younger when you leave age ambiguous, so you anchor it up front and reinforce it in the negative line.
If you are building a recurring character, this is where identity tokens, a trained LoRA trigger word, or an IPAdapter reference do their work. For the full identity toolkit see character consistency techniques and IPAdapter for character consistency.
Slot 2: action or pose
What is the subject doing? standing, contrapposto, looking at viewer, one hand on hip. Pose is where most prompts quietly fail because text alone is a weak lever for body geometry. If a pose keeps collapsing, you stop fighting the prompt and reach for pose control. The full pose library lives in NSFW AI pose prompts, and the locking method is in OpenPose pose control.
Slot 3: outfit and clothing
Describe garment, material, and state of coverage in tasteful tag terms. silk robe, partially open, lace trim. Material words (silk, latex, cotton) do more work than color words because they drive how light behaves. The complete phrasing set is in NSFW AI outfit prompts.
Slot 4: setting and background
Where are they? bedroom, soft bedsheets, blurred background. A clean background keeps attention on the subject and reduces artifact risk. Build settings with NSFW AI setting prompts.
Slot 5: lighting
Lighting is the single biggest lever on mood and realism. soft studio lighting, rim light, golden hour. Hard light sculpts, soft light flatters. Deeper treatment in NSFW AI lighting prompts.
Slot 6: camera, shot type, lens
Shot type and focal length decide framing and depth. medium shot, 85mm, shallow depth of field, bokeh. The full set of angles and lenses is in NSFW AI camera angle prompts.
Slot 7: art style or medium
photorealistic versus anime screencap versus oil painting. This is the slot where SDXL, Pony, Illustrious, and Flux diverge most, covered below.
Slot 8: quality and score tags
Finishing tokens. On Pony and Illustrious these are the famous score_9, score_8_up, score_7_up chains. On SDXL realistic checkpoints it is more like best quality, highly detailed, 8k. On Flux you can usually drop them entirely.
Slot 9: the negative line
Always last, always present. Your negative carries the baseline safety tokens plus quality cleanup. Build a real one with negative prompts that work.
Per-model differences that actually matter
The formula is universal. The dialect is not. Here is how the same intent gets phrased across the four families.
| Model family | Prompt language | Quality or score tags | Best for |
|---|---|---|---|
| SD 1.5 / SDXL realistic | Tag plus light natural language | best quality, highly detailed | Photoreal portraits |
| Pony Diffusion | Booru tags, heavy | score_9, score_8_up, score_7_up | Stylized and concept range |
| Illustrious XL | Booru tags, danbooru aligned | masterpiece, best quality | Anime and illustration |
| Flux | Plain natural language sentences | usually none | Coherent realism, hands |
SDXL: natural language friendly tags
SDXL realistic checkpoints understand short descriptive phrases mixed with tags. You can write almost conversationally and still get clean results. Pick your base from best NSFW checkpoints.
1woman, adult, 27 years old, confident standing pose, looking at viewer,
silk robe partially open, modern bedroom, soft window light, rim light,
medium shot, 85mm, shallow depth of field, photorealistic, best quality, highly detailed
Negative: child, minor, underage, loli, shota, lowres, bad anatomy, bad hands,
extra fingers, deformed, watermark, text, jpeg artifactsPony Diffusion: score tags lead the line
Pony and its derivatives were trained with score prefixes. They genuinely improve output, so they go near the front, then your booru tags. Full setup in Pony Diffusion guide and tag reference in danbooru tags for NSFW AI.
score_9, score_8_up, score_7_up, 1girl, adult, mature female, solo,
standing, looking at viewer, contrapposto, silk robe, partially open,
bedroom, soft lighting, rim light, medium shot, depth of field
Negative: score_1, score_2, child, minor, underage, loli, shota,
lowres, bad anatomy, bad hands, extra digits, deformed, watermark, signatureIllustrious XL: danbooru aligned tags
Illustrious leans on clean danbooru tagging and rewards masterpiece, best quality over score chains. Read Illustrious XL guide and how to use Illustrious models.
masterpiece, best quality, 1girl, adult, mature female, solo, standing,
looking at viewer, off-shoulder dress, indoor, window light, cinematic lighting,
cowboy shot, depth of field, anime style
Negative: child, minor, underage, loli, shota, worst quality, low quality,
bad anatomy, bad hands, extra digits, jpeg artifacts, watermarkFlux: write a sentence, not a tag soup
Flux is the outlier. It parses natural language sentences far better than tag strings and produces excellent hands and coherent scenes. Drop the score tags and just describe. See Flux NSFW guide.
A photorealistic medium shot of an adult woman, around 28, standing confidently
by a window in a modern bedroom. She wears a partially open silk robe. Soft morning
light comes through the window with a gentle rim light on her shoulder.
Shot on 85mm with a shallow depth of field and creamy bokeh.
Negative: child, minor, underage, loli, shota, deformed, extra fingers,
watermark, text, low qualityA worked photoreal example, slot by slot
Let us build one cleanly so you can see the order pay off.
1woman, adult, 30 years old,
sitting on edge of bed, leaning back on hands, looking at viewer,
lace bodysuit, sheer fabric,
dim bedroom, blurred background,
low key lighting, warm rim light,
close-up to medium shot, 50mm, shallow depth of field, bokeh,
photorealistic, film grain,
best quality, highly detailed, 8k
Negative: child, minor, underage, loli, shota, lowres, bad anatomy,
bad hands, extra fingers, fused fingers, deformed, mutated, watermark,
signature, text, jpeg artifacts, blurry faceNotice the adult tag sits at slot one and the safety tokens lead the negative. That is the spine of every responsible NSFW prompt. For realism specific tactics see how to make realistic AI images.
Ready to iterate fast? Open the generator and change one slot at a time. Single variable testing is the only way to learn what each slot actually controls.
Prompt weighting: the advanced lever
Once the order is right, you fine tune emphasis with weights. In most A1111 style syntax, (token:1.3) strengthens and (token:0.7) weakens. Use it sparingly. Over weighting one token at the expense of everything else is the most common cause of fried, oversaturated output. The full discipline is in NSFW AI prompt weighting.
1woman, adult, (confident expression:1.2), standing, (silk robe:1.1),
bedroom, (soft rim light:1.2), 85mm, photorealistic, best quality
Negative: child, minor, underage, loli, shota, bad hands, deformed, watermarkCommon formula mistakes
| Mistake | Symptom | Fix |
|---|---|---|
| Style token first | Generic, off concept | Put subject first |
| No adult tag | Drift and safety risk | Anchor adult up front and in negative |
| Score tags on SDXL | No effect or worse | Use best quality on non Pony models |
| Tag soup on Flux | Incoherent scene | Write a sentence |
| Everything weighted | Fried image | Weight one or two tokens max |
| Empty negative | Artifacts and bad hands | Always run a full negative line |
How to debug a prompt the formula way
When an image disappoints, resist the urge to rewrite the whole thing. The formula gives you a diagnostic ladder. Work it top to bottom.
First, check slot one. Is the subject count clear? A stray plural can spawn a second figure. Is the adult tag present and reinforced in the negative? If the face reads too young, that is always the first fix.
Second, isolate the pose. Render with the pose tokens removed entirely and see if the body cleans up. If it does, your pose phrasing is fighting the model and you should move to OpenPose or ControlNet rather than piling on more pose words. Heavy pose stacking is a classic cause of melted limbs.
Third, look at clothing conflicts. Contradictory garment tags (for example two different dress types) confuse the model and produce hybrid fabric blobs. Pick one garment and one state of coverage.
Fourth, simplify the background. A busy setting steals denoising budget from the subject. Add blurred background or simple background and watch the subject sharpen.
Fifth, audit weights. If the image looks fried, oversaturated, or has burned highlights, you almost certainly over weighted a token. Drop everything back toward 1.0 and reintroduce emphasis one token at a time.
This ladder works because each slot maps to a different failure mode. Random rewriting does not teach you anything. Slot by slot debugging turns every bad render into a lesson about which lever does what.
Sampler and settings: the formula is only half the render
A perfect prompt still needs sane generation settings. The prompt decides what to draw, the sampler and steps decide how cleanly it gets drawn. For most NSFW work, a sampler like DPM++ 2M Karras or Euler a at 25 to 35 steps is a safe default. CFG around 5 to 7 keeps the model faithful to the prompt without frying it. Push CFG too high and you reintroduce the same burned look that over weighting causes.
Resolution should match what the model was trained on. SDXL, Pony, and Illustrious want roughly 1024 by 1024 or equivalent aspect ratios. Generating far below native resolution produces soft, low detail results that no prompt can rescue. When you need a larger final, generate at native size then upscale, rather than generating huge in one pass. For the deeper pipeline including upscaling and cleanup, see the NSFW photo editing workflow.
If you run your renders in ComfyUI, the same formula maps cleanly onto nodes: a positive conditioning node for your slots and a negative conditioning node for your safety line. The ComfyUI for NSFW guide shows the graph.
LoRAs slot into the formula too
A LoRA is a small add on that teaches the base model a specific concept, character, style, or detail. In the formula, the LoRA trigger word usually rides in slot one or seven depending on what it does. Identity LoRAs go up front with the subject. Style LoRAs sit near the art style slot. Keep LoRA strength moderate, often 0.6 to 0.9, because stacking several at full strength overwhelms the base model and produces incoherent output. Browse options in best NSFW LoRAs.
1woman, adult, <lora:examplecharacter:0.8>, standing, looking at viewer,
satin slip dress, studio backdrop, soft key light, rim light,
medium shot, 85mm, photorealistic, best quality, highly detailed
Negative: child, minor, underage, loli, shota, bad anatomy, bad hands,
extra fingers, deformed, watermark, text, low quality
Multiple characters need regional control
The single subject formula breaks down the moment you want two distinct adult characters in one frame, because the model blends their attributes. The fix is not a longer prompt, it is regional prompting, where you assign different prompt blocks to different areas of the canvas. See Regional Prompter for multiple characters and combine it with ControlNet for pose accuracy across both figures. Each character still follows the nine slot order inside their own region, and each region keeps the adult tag.
Build your prompt library, not just one prompt
The creators who win treat prompts as reusable modules. They keep a swipe file: a few subject blocks, a dozen pose blocks, a lighting set, a camera set, and a battle tested negative line. Then they assemble like building blocks. That is exactly why this cluster is split by slot, so each piece is a module you can drop into the formula. Skim more starting points in NSFW AI prompt examples and the best NSFW prompt generators. For uncensored base options to run the formula on, see best uncensored AI image generators and best NSFW AI image generators.
The payoff of modular prompting is speed and consistency. Once your swipe file is built, producing a new on brand image is a matter of swapping one or two blocks rather than writing from scratch and hoping. Your safety negative line never changes, so it becomes muscle memory: every assembly starts and ends with it locked in.
When you are ready to assemble and render, the generator is here. Keep every subject adult, fictional, and AI-generated, never a real or identifiable person, and keep your safety negatives locked on every single render.
Frequently asked questions
What is the basic NSFW AI prompt formula?
Order your prompt as subject plus adult tag, then action or pose, outfit, setting, lighting, camera, art style, quality or score tags, and a negative line last. Earlier tokens carry more weight, so put what matters most first. Keep baseline safety negatives like child, minor, underage, loli, shota on every render.
Does prompt order really change the image?
Yes. Diffusion models weight earlier tokens more heavily than later ones, so moving the subject to the front versus the style to the front produces different compositions. Front load the who and what, back load the polish like quality tags. When output looks wrong, walk the slots in order to find which one the model ignored.
Why do Pony models use score_9 and SDXL does not?
Pony Diffusion was trained with score prefixes baked into its dataset, so tokens like score_9 and score_8_up genuinely improve output. SDXL realistic checkpoints were not trained that way, so those tags do nothing useful there. On SDXL use best quality and highly detailed instead, and on Flux you can usually drop quality tags entirely.
How is Flux prompting different?
Flux parses natural language sentences far better than tag strings. Instead of a comma separated tag soup, describe the scene in plain English: who the adult subject is, what they wear, the lighting, and the camera. Skip score tags. This gives Flux its strong hands and coherent compositions. Keep your safety negatives on regardless.
Where should the adult age tag go?
Put an explicit adult marker like adult, mature female, or 27 years old in slot one right next to the subject, and reinforce it in the negative line with child, minor, underage, loli, shota. Models drift younger when age is ambiguous, so anchoring it up front is both a quality and a safety decision on every NSFW render.
Do I need every slot in the formula?
No. The nine slots are a checklist, not a requirement. Many strong prompts skip the art style or quality slot. The value of the formula is consistency and troubleshooting: when something looks off, you check slots in order to find the weak link. The only slot you never skip is the negative line with safety tokens.
How do I lock a pose the prompt keeps ignoring?
Text is a weak lever for body geometry, so when a pose collapses repeatedly, switch from describing it to controlling it with OpenPose or ControlNet. You supply a skeleton or reference and the model follows it. Our pose and OpenPose guides cover the full workflow. Prompt phrasing gets you close, pose control locks it in.
Can I prompt for a real person to make NSFW images?
No. Generating NSFW images of a real identifiable person without consent is unethical and illegal in many places under deepfake and non consensual intimate image laws. Every subject in these guides is an adult, fictional, AI-generated character. Use original or owned personas only, and always keep minor and likeness safety negatives active.
