SeaArt Prompts: How to Write Them for Any Model or Style
A good SeaArt prompt names the subject first, adds setting, light and lens, puts any LoRA trigger words near the start and finishes with one or two style words. Pair it with a short negative prompt, match the prompt style to the model family you have selected, and reproduce an example image’s settings before changing anything.
SeaArt (seaart.ai) gives you a very large menu of community models and styles, which is both its strength and the reason prompts that work in one place fail in another. The same sentence behaves differently on an anime SDXL checkpoint, a photoreal checkpoint and a Flux-based model. This guide gives you a prompt structure that transfers, shows how to adapt it to each family, and explains how to borrow settings from images already on the site.
Start from an image, not a blank box
Every public image on SeaArt carries the data that made it: model, LoRAs, prompt, negative prompt, sampler, steps, CFG, seed and size. The fastest route to a working prompt is to find an image close to what you want, open its details and use the option that loads those settings into the generator. Run it unchanged once. If your result matches, you have a proven starting point, and every later difference is yours to learn from. Only then start editing the prompt.
Note the model family while you are there. SeaArt labels the base model on each checkpoint and LoRA (SD 1.5, SDXL, Flux and others), and LoRAs only work with checkpoints from the same family.
A prompt structure that transfers
Write in this order and most models will follow you:
- Trigger words. If a LoRA lists one, put it first, spelt exactly as the card shows.
- Subject and action. "a tired courier resting on a bicycle". One subject, one action, in plain words.
- Setting. "outside a late-night bakery, wet pavement".
- Light and time. "warm shop light spilling onto the street, blue hour".
- Camera. "35mm lens, eye level, shallow depth of field". This does more for realism than any quality word.
- Style, briefly. One or two terms: "documentary photograph" or "flat vector illustration". Avoid stacking ten.
Keep the whole thing under roughly 75 tokens on SD 1.5 and SDXL models. Longer prompts are processed in chunks and the later parts lose influence. If you need more, cut adjectives before you cut nouns.
Adapting to the model family
- SDXL and SD 1.5 checkpoints. Comma-separated phrases work well. Emphasis syntax such as
(red umbrella:1.3)is supported in most interfaces of this kind; use it on at most two elements. A negative prompt matters here. - Anime and illustration checkpoints. Many are trained on tag-style captions. Short tags separated by commas ("1girl, school uniform, rooftop, sunset") often beat sentences. The model card usually says which captioning style it expects.
- Flux-based models. Prefer full sentences, respond poorly to quality padding, and often ignore or lack a negative prompt. Describe what you want to see, not what you want to avoid.
- Photoreal checkpoints. Lens, light and film words are the levers. "Kodak-style colour" is a style; a brand logo in the frame is not something to prompt for.
Negative prompts and weighting
A negative prompt tells the model what to steer away from. A short general list covers most SDXL work: "blurry, low quality, deformed hands, extra fingers, text, watermark, cropped". Add a specific term only when a specific fault keeps appearing, such as "second person" when a single-subject prompt produces two. Very long negatives flatten detail, so trim them when results start looking waxy.
Weighting goes the other way. If the model keeps dropping an element, raise it to 1.2 or 1.3 with parentheses. If one element is taking over, lower it to 0.8. Change one weight at a time with a fixed seed so you can see the effect.
The prompt helper and when to use it
SeaArt, like most community platforms, includes tools that expand or translate a short prompt into a longer one. The exact name and placement change with updates, so look for an enhance or optimise option near the prompt box. It is useful for discovering vocabulary you had not thought of, and for non-native English writers who want idiomatic phrasing. It is less useful when you already know what you want, because it tends to add generic quality words and can shift the subject. Read what it produces, keep the phrases that add information and delete the rest.
Settings at a glance
| Setting | What it does | Start with |
|---|---|---|
| Model | Base look and family | One whose examples match your goal |
| LoRA weight | Strength of a style or subject add-on | Card value, else 0.7 |
| Prompt length | How much the model reads with full weight | Under 75 tokens on SD-family models |
| Emphasis | Raise or lower one element | (element:1.2) on one or two words |
| Negative prompt | What to avoid | blurry, low quality, deformed hands, text, watermark |
| CFG | Prompt adherence | 5 to 7 on SDXL, low on Flux |
| Seed | Reproducibility | Fixed while editing a prompt |
For a repeatable face or outfit across many images, see SeaArt consistent characters. For what sampler, steps and CFG do, see the Stable Diffusion prompts guide.
Try it
Model: a photoreal SDXL checkpoint
Prompt: a tired courier resting on a bicycle outside a late-night bakery, wet pavement, warm shop light spilling onto the street, blue hour, 35mm lens, eye level, shallow depth of field, documentary photograph
Negative prompt: blurry, low quality, deformed hands, extra fingers, text, watermark, cropped
Size: 1024x1024. Steps: 28. CFG: 6. Seed: fixed after the first run
Once you have a keeper, swap "documentary photograph" for "flat vector illustration" on the same seed. The composition will mostly hold while the style changes, which shows you how much of the image the style word controls on that model.