Illustrious prompt guide, tested
Write Illustrious prompts as comma-separated Danbooru tags in roughly this order: who (solo, mature female), what they look like, what they wear, what they’re doing, the framing, the place, and the light. We tested nine popular rules on WAI-Illustrious v17, each on the same three seeds. Most barely changed the image. Three things did: the rating tag, the adult tag, and the actual content tags.
Unless a figure says otherwise, every render below uses this base prompt on seeds 101, 202 and 303, run on SirenRender’s free anime tile (WAI-Illustrious-SDXL v17 with the DMD2 LoRA, 8 steps, CFG 1). The tile adds masterpiece, best quality on its own.
1. Start with a rating tag
Illustrious models were trained with rating tags, and the WAI author lists four: general, sensitive, nsfw and explicit. This is the strongest single tag we tested. With general the same beach outfit stays modest; sensitive pushes it toward fan service; leaving the tag out lands in between.
Prompt for this test: [rating], solo, mature female, long black hair, brown eyes, white sundress, straw hat, smile, standing, beach, ocean, summer, blue sky, cowboy shot, looking at viewer
2. Say the character is an adult
WAI’s default character skews young. Its author’s fix, from the model page: “try adding (aged up:1.0–2.0) or (mature female:1.0–2.0) at the beginning.” In our test, 1girl on its own gave the youngest-looking faces. Plain woman kept the same single-character framing as 1girl, and mature female (with or without aged up) gave clearly adult women. Every Illustrious prompt in this library uses mature female or mature male.
3. One set of quality tags is enough
The WAI model page recommends masterpiece, best quality, amazing quality and warns: “Please do not add too many quality and aesthetic-related tags, nor overly long negative prompts, as this will actually reduce image quality and make it more blurry.” On the DMD2 setup we saw no visible gain from more. Adding amazing quality, or a whole stack of very aesthetic, absurdres, highres, newest, ultra-detailed, gave nearly the same picture as the tile’s own two tags.
4. Tag order matters less than you’d think
We shuffled the base prompt into a random order. Same outfit, same street, same light, and on seed 101 almost the same frame. Keep a readable order because it’s easier to edit, not because the model needs it.
5. Plain sentences work too
WAI v17 follows natural language well enough to mix it with tags. Written as one sentence, the base prompt gave the same woman, sweater and street. The hybrid version added a sentence about buckets of tulips, and the tulips showed up.
6. Underscores or spaces: either works
Danbooru writes tags with underscores (long_hair), most prompts use spaces. Both produced the same character and scene.
7. Weights nudge, they don’t transform
ComfyUI, which SirenRender runs on, reads (tag:1.3) weights (the parser is in sd1_clip.py). Weighting the hair and the street to 1.3 changed details, not the picture. Reach for a weight when a tag is being ignored, not by default.
8. BREAK only means something in A1111 and Forge
BREAK is an AUTOMATIC1111 feature that pads the current 75-token chunk and starts a new one. ComfyUI’s text encoder has no BREAK handling, so on ComfyUI-based generators it’s just one more word. In our test it moved small details only.
9. Negative prompts
At CFG 1, which is how fast DMD2 and Lightning setups run, the negative prompt isn’t evaluated at all. The negative prompts guide has the details and what to write instead.
A template to start from
Every slot has a visual reference page: framing and angles, poses, expressions, hairstyles, hair colors, backgrounds, lighting and art styles. For finished prompts, see the Illustrious prompt library.
What we didn’t find
Three seeds per rule is enough to see a strong effect, not to rule out a subtle one. Every finding above is from WAI-Illustrious v17 at 8 steps and CFG 1. At the model’s own recommended settings (25–40 steps, CFG 5–7, Euler a) quality tags and negatives have more room to act.