Danbooru Tags for NSFW Anime AI: Order, Adult Cues and a Cheat Sheet

If you write prompts for an uncensored anime model, you are writing Danbooru tags whether you call them that or not: 1girl, cowboy shot, from above. This guide is about writing them well, on two models that read them differently.

Danbooru tags work best in a fixed order (quality, rating, style, head count, adult cues, appearance, wardrobe, pose, camera, scene and light) and only when the words really exist as tags. In 42 test images on 6 October 2026, an invented phrase like "desert motel at blue hour" was dropped in all six Prefect Pony XL runs, while the real tags dusk, twilight, gradient sky, motel put a dusk sky in both runs that used them. Pony keeps the character and needs its score string. Neta Lumina follows more of the scene, ignores Danbooru rating words and drifts toward a semi-realistic painting unless you add anime illustration.

We ran everything through Neta Lumina and Prefect Pony XL LoRA, both uncensored with no content filter. Every prompt printed here describes clothed adults and is suggestive at most. The platform does not provide prompts for stronger content; we describe how far each model goes with our level names instead.

At a glance

At those prices, all 42 test images in this article cost about $0.82.

The order that worked on both models

Both models learned Danbooru-style tags, so one skeleton serves both. Neta's prompt book puts quality tags last; Pony's convention puts score, source and rating first. We used one order for every prompt here, quality first on both, and did not test moving it to the end.

Here is the full Pony prompt we sent for the cover scene:

score_9, score_8_up, score_7_up, source_anime, rating_safe, 1girl, solo, mature female, adult woman, 34 years old, mature face, smile lines, long dark wavy hair, brown skin, astronomer, white shirt with rolled-up sleeves and an unbuttoned collar, high-waisted dark trousers, leaning against a large brass telescope, looking at viewer, slight smile, observatory dome at night, dome slit open to the stars, red night lamps, cold blue moonlight, wide shot

For Neta, swap the first five tags for masterpiece, best quality, anime illustration and keep the rest. We also ran the list without anime illustration on Neta, to see how the two models read identical tags:

Left: Prefect Pony XL LoRA. Right: Neta Lumina. Identical tags after each model's own opening, no style tag on Neta, seed 11, 16:9, text to image.

The difference is the clearest lesson in this guide. Pony drew a more polished character and cut the scene: no dome, no red lamps, and a waist-up shot instead of the wide shot we asked for. Neta built the observatory, the red lamps and the wide frame, but rendered it as a semi-realistic painting. Pony ranks the character first. Neta reads further down the list. On Pony, write the scene in words it knows (next section); on Neta, add a style tag.

Use words that exist as Danbooru tags

The tags a model knows are the words that appeared in its training captions. A phrase you invent is read as loose text, and on Pony it is the first thing to go. We tested it with one adult woman sitting on a vintage car, changing only the scene words:

Prefect Pony XL LoRA, 3:2. Left column: desert motel at blue hour, parking lot, distant mesas, pink and blue sky. Right column: dusk, twilight, gradient sky, purple sky, motel, building, parking lot, desert, mesa. Both columns include detailed background, scenery. Top seed 61, bottom seed 62.

With the invented phrase, all six runs we made (with and without detailed background, scenery, and with the scene moved to the front) came out in bright daylight with no motel. With the real tags, both runs got a purple dusk sky, and one got a long motel building. The prompt that worked:

score_9, score_8_up, score_7_up, source_anime, rating_safe, 1girl, solo, mature female, adult woman, 33 years old, mature face, curly red hair, freckles, cropped denim jacket over a white tank top, high-waisted jeans, cowboy boots, sitting on the hood of a vintage car, looking at viewer, smirk, detailed background, scenery, dusk, twilight, gradient sky, purple sky, motel, building, parking lot, desert, mesa

detailed background, scenery is still worth keeping. On its own it did not bring the motel back here, but in our example runs a vault scene drew no vault until we added it together with concrete nouns (inside a bank vault, giant round vault door). Name objects, not moods.

Spaces or underscores?

Danbooru itself writes long_hair, and some guides insist on underscores. Users in the comments under one of those guides say spaces worked better for them. We ran the same prompt both ways on both models:

Top: Prefect Pony XL LoRA. Bottom: Neta Lumina. Left: tags with spaces. Right: the same tags joined with underscores. Seed 21, 1:1.

Neta did not care. On this seed, Pony lost the grey hair and the age when we wrote short_grey_hair and 40_years_old, which are not real tags in either form. Our rule: write spaces, and keep underscores only in tokens that are defined that way, such as score_8_up, source_anime, rating_safe and LoRA trigger words like end_inkpunk_style.

Quality, score and rating tags

Pony's score string. Pony Diffusion's author explained on Civitai that score tags come from roughly 20,000 hand-ranked training images, and that the model learned the whole string score_9, score_8_up, score_7_up as one signal for "good looking". That is why score_9 alone is weak. The Pony V6 card lists the full ladder down to score_4_up; the first three are the common short form and what we used. They do nothing useful on Neta. Pony also has source tags (source_anime, source_cartoon, source_furry, source_pony) and three rating tags: rating_safe, rating_questionable and rating_explicit.

Neta's opening. Neta documents no rating tags at all. Its published tag examples, and every Neta prompt here, open with masterpiece, best quality.

Do rating tags change the picture? We kept one clothed, suggestive prompt and swapped only the rating word:

Top: Prefect Pony XL LoRA with rating_safe, with no rating tag, and with rating_questionable. Bottom: Neta Lumina with no rating word, with safe, and with sensitive. Same prompt otherwise, seed 71, 2:3.

score_9, score_8_up, score_7_up, source_anime, rating_safe, 1girl, solo, mature female, adult woman, 30 years old, mature face, long platinum hair, oversized cream knit sweater slipping off one shoulder, black shorts, sitting on a window seat, legs crossed, holding a mug, looking at viewer, soft smile, rainy afternoon, cozy loft apartment, warm lamp light, cowboy shot

On Pony at seed 71, rating_safe gave a looser sweater and a higher neckline, and dropping it or switching to rating_questionable made the figure fuller and the neckline lower. On a second seed the three Pony versions were much closer to each other. In all six Pony runs the clothing the prompt described stayed on. On Neta the three images were nearly identical.

So the description sets the level and the rating tag only nudges it. Our content tests on 5 October point the same way: the Pony ladder carried rating_questionable on its first rung and rating_explicit above it, the Neta ladder carried no rating word at all, and the two scored almost the same. In those Freedom runs, both delivered every run from L1 Suggestive (lingerie, a slipping strap) through L2 Partial and L3 Strong to L4 Intense (an intimate scene between two adults). At L5 Extreme (consensual restraint between adults, or cinematic gore) Neta delivered every run, while Pony softened all three restraint runs and two of four gore runs. That is why the Freedom Scores are 100 and 92.5.

Keep every character clearly adult

1girl and 1boy are head-count tags; they say how many people are in the frame, not how old anyone is. On Danbooru, mature_female means a woman who looks middle-aged, which is why it pulls faces older. Age has to be something the model can see, so we stack cues:

• mature female or mature male, plus adult woman or adult man
• an age number, 30 years old or older
• one or two visible features: mature face, smile lines, sharp cheekbones, stubble
• an adult role and setting: tattoo artist, astronomer, bartender, a hotel bar rather than a school

Leave out anything youth-coded, such as school uniforms, classrooms, "cute" or "youthful". Our adult anime style guide has more on faces and proportions.

What we rejected. For the Pony example of a tattoo artist, the first round used mature female, adult woman, 32 years old with an ink LoRA at 0.8, and all three seeds read too young. The kept version raised the age to 35, added mature face, sharp cheekbones, smile lines, lowered the LoRA to 0.5 and dropped two ink tags. We changed several things at once, so we cannot say which one fixed it; the age cues were the deliberate part.

Prefect Pony XL LoRA with the InkPunk LoRA at 0.5, seed 107, 2:3. The kept version, after the first round read too young.

The two models lean opposite ways. Pony drifted younger when we gave it sentences: a keeper "in his mid-forties" came back as a younger man in both seeds (see the lighthouse test below). Neta overshot: the same mature face, smile lines on Neta gave one 34-year-old a heavily lined face. On Pony, stack the cues; on Neta, an age number and one cue are usually enough. On both, look at every face before you keep an image.

Two characters: one phrase per person, actions as tags

With two people, attributes bleed: her colour ends up on his shirt. Put both head-count tags first, then give each person one phrase that starts with "she" or "he":

Neta Lumina, seed 231, 16:9. Head-count tags first, then one phrase per person.

masterpiece, best quality, anime illustration, 1girl, 1boy, hetero, mature female, mature male, adult couple in their thirties, she has long wavy platinum hair and wears a deep crimson satin slip dress with a black lace shawl, he has slicked-back black hair and wears an open-collar black shirt under a long dark coat, she sits on the rim of a stone fountain, he stands over her with one hand lifting her chin, faces almost touching, gothic cloister at night, huge blood-red moon, ravens on the arches, red and black palette, wide shot

What about the action between them? We expected a sentence to beat tags here, and on Neta it did not. We tested a couple in a sleeper train compartment, once with the action as tags (face to face, adjusting necktie) and once with the action moved into sentences.

Neta Lumina, 3:2. Left column: all tags, with face to face, adjusting necktie. Right column: the same skeleton with the action in sentences. Top seed 41, bottom seed 42.

The tag version held the tie action in both seeds. The sentence version picked up the half smile we described, but the action itself got vague. If Danbooru has a tag for the action, use the tag, and save sentences for expression and mood.

Camera and light tags

Framing tags are some of the most reliable Danbooru tags. We kept one character and seed on Pony and changed only the camera tag:

Prefect Pony XL LoRA, seed 31, 2:3. Top: upper body, cowboy shot, full body. Bottom: from below, from above, dutch angle.

The same trial also showed scene drift: we asked for a ferry deck and got a lantern-lit waterfront in all six frames. On Neta, camera tags in a tall frame worked well in our example runs: an earlier wording of a leaping duelist without pov and fisheye came out side-on, and pov, extreme foreshortening, fisheye aimed the leap at the viewer.

For light, use words that exist as tags: night, dusk, sunset, moonlight, backlighting, rim light, lamp. The dusk sky in the motel test came from exactly this kind of word.

Style tags and LoRA trigger words

Neta takes its style from a tag. Without one, our Neta images came back as glossy semi-realistic paintings. One tag fixed it:

Neta Lumina, seed 21, 1:1. Left: the tag list alone. Right: the same list with anime illustration added after the quality tags.

One phrase to avoid on Neta: "key visual". In our example runs, prompts with it came back with trading-card frames, pseudo-text and a rarity-style badge, so we dropped it.

Pony takes its style from a LoRA, and the trigger goes first. Each style LoRA we used wanted its own trigger at the very start: vztdzsxx for Flat Vector, end_inkpunk_style for InkPunk, end_artdeco_inspo for Art Deco. Tags after the trigger can push the look further. Our first Flat Vector courier read as ordinary anime; standing her up and adding bold outlines, limited palette, simple shapes got the poster look:

Prefect Pony XL LoRA with the Flat Vector LoRA at 1, trigger vztdzsxx first, seed 137, 9:16.

Use only LoRAs trained for Pony-family SDXL; one for another base model loads without an error and changes nothing. Scale, stacking and the silent wrong-base check are covered in our Pony XL LoRA guide. Neta has no LoRA slot, so its style comes from tags only.

Tags, sentences or both

People often ask how to turn Danbooru tags into natural language. On these two models the answer depends on which one you use. We sent the same three sentences, a lighthouse keeper in his mid-forties with a salt-and-pepper beard, a cable-knit sweater, an open yellow raincoat and a storm lantern, to both:

The same three sentences. Left pair: Prefect Pony XL LoRA, seeds 51 and 52. Right pair: Neta Lumina, seeds 51 and 52. 3:4.

Pony picked out the keywords it knows (yellow raincoat, lighthouse, rain, beard) and dropped the sweater, the lantern and the age in both seeds. Neta drew the sweater, the raincoat, the lantern and the grey beard in both, and made him older than we asked.

Cheat sheet

A template to copy:

[quality], [source and rating on Pony], [style or trigger], [head count], [adult cues], [hair, skin], [one phrase per garment], [pose, gaze, expression], [camera], [scene], [light]

When a tag does not land

Which model for which habit

For more on each model, read our Neta Lumina test and the Pony XL LoRA guide linked above.

FAQ

What does score9, score8up, score7_up mean?

They are Pony's quality tags. Pony's author ranked about 20,000 training images by hand, and the model learned the whole string as a signal for its best-looking images. Put it first on Pony models; score_9 alone is weak, because the model learned the whole string rather than its parts. It has no use on Neta Lumina.

Can rating tags control how NSFW the image gets?

Only a little. On Pony, removing rating_safe or using rating_questionable made one clothed scene a little more revealing, but the clothing the prompt described stayed on. On Neta, safe and sensitive changed nothing. Your description sets the level.

Should Danbooru tags use underscores or spaces?

Spaces. Neta read both the same way. Pony lost details when we joined invented phrases with underscores. Keep underscores only in tokens defined that way, such as score_8_up, rating_safe and LoRA triggers.

Does 1girl mean a young character?

No. 1girl and 1boy count people. Age comes from cues you add: mature female, adult woman, an age of 30 or more, mature face. Check every face and reroll anything that reads young.

Can I turn Danbooru tags into natural language?

On Neta Lumina, yes. It followed three plain sentences closely in our test. On Pony, sentences lost the sweater, the lantern and the age, so keep tags there.

Does Pony XL need a negative prompt?

No. The Pony V6 card says it rarely needs one, and neither endpoint here has the field. Write the opposite as a positive tag (simple background, solo, fully clothed) and let the description steer.

Where can I find a full list of Danbooru tags?

The tag groups page on Danbooru's wiki lists hair, clothing, composition, angles and places, and it is the list the community threads we read point to. Check that a tag exists there before you build a prompt around it.