From Danbooru Tags to Sentences: Rewriting Anime Prompts for Seedream, Wan and Nano Banana

Booru tags, the Danbooru-style labels anime image boards attach to every picture, look like long_hair, red_eyes, looking_at_viewer. Models trained on those captions, such as NovelAI’s anime models and many Stable Diffusion fine-tunes, learned each tag as a switch, which is why a comma-separated tag prompt works so well on them. NovelAI’s documentation says as much: its models recognise tags they were trained on better than loose wording.

The image models in SpicyHub AI’s Image Lab learned differently. Seedream 5.0 (offered here as Spicy Imagine Pro), Wan 2.7 Image and Nano Banana 2.1 are general models that you prompt in ordinary language. Hand them a tag list and they still pick out familiar words, but the grammar that ties the words together is gone: whose hair is silver, who holds the umbrella, where the light comes from. None of them has a weighting syntax either, so brackets and numbers such as (red eyes:1.3) or {{red eyes}} are just extra characters.

Even NovelAI now describes its newer models, from V4 on, as understanding natural language. The fix is not to write more. It is to write the same information as two or three plain sentences, in an order the model can follow.

Write the layers in this order and keep each one short:

1. Subject: how many people, that they are adults, an age range and a build. 2. Features: hair, eyes, skin and anything else that identifies the character. 3. Clothing: garments, colours and materials. 4. Action and expression: the pose, what the hands are doing, the mood. 5. Framing: how much of the figure is in shot, and the camera angle. 6. Setting: the place and the time of day. 7. Light: where it comes from and what colour it is. 8. Style: the medium and finish, such as a cel-shaded anime illustration with clean line art.

Layers 1 to 4 usually make the first sentence, 5 to 7 the second, and the style closes the prompt. Put the subject first so nothing else competes with it, and drop the underscores and commas along with the tag format.

Count and subject 1girl, solo → one adult woman, alone in the frame 1boy, male focus → one adult man as the clear subject 2girls → two adult women, then one sentence for each of them mature female → an adult woman in her thirties (state the age range outright)

Hair long hair, silver hair, ponytail → long silver hair tied in a high ponytail ahoge → a single strand of hair sticking up at the crown hair over one eye → her fringe falls over her right eye

Eyes and face red eyes, slit pupils → red eyes with narrow, cat-like pupils smile, closed mouth → a small closed-mouth smile blush → a light flush across her cheeks

Clothing black jacket, white shirt, necktie → a black tailored jacket over a white shirt and a loosened dark tie detached sleeves → separate sleeves that begin above the elbow hooded cloak, hood up → a long hooded cloak with the hood raised

Pose and action standing, arms crossed → standing with her arms folded looking at viewer → looking straight at the camera looking back → glancing back over her shoulder

Framing and camera portrait, close-up → a head-and-shoulders close-up upper body → framed from the waist up cowboy shot → framed from mid-thigh up full body → the whole figure in shot, head to feet from below → seen from a low angle, looking up

Background and light outdoors, night, city lights → outside at night, with city lights blurred behind her simple background, white background → a plain white background backlighting → lit from behind, with a bright rim of light around her hair sunset → warm orange sunset light from the left

Style anime coloring → flat, cel-shaded colour sketch, monochrome → a loose pencil sketch in greyscale watercolor (medium) → soft watercolour washes on textured paper

Quality and meta tags: delete them masterpiece, best quality, highres, absurdres, very aesthetic, artist names → nothing. Describe the look you want instead of borrowing another artist’s name or a quality label.

Booru count tags use the same words for every female or male character, whatever their age, and anime styling can make a grown character look under 18 by accident. A model that reads sentences takes your words literally, so let the first sentence settle the question: an adult woman of about twenty-eight, a man in his mid-forties, adult proportions, an adult setting such as an office, a bar or a flat of their own.

Leave out every tag that would make a character read as under 18, whether it describes the body, the clothes or the setting. Such tags don’t belong in a prompt on SpicyHub AI in any art style, and the terms say so. Use original characters too: existing anime characters belong to their rights holders.

1. A character portrait on Nano Banana 2.1 (2K, 12 credits) Tags: mature female, solo, shoulder-length black hair, amber eyes, grey coat, red scarf, upper body, looking at viewer, smile, snow, night, street lamp, masterpiece, best quality Sentences: An adult woman aged about thirty-two, with shoulder-length black hair and amber eyes, smiles at the camera in a grey wool coat and a red scarf. She stands under a street lamp on a snowy night, framed from the waist up, with soft flakes catching the light. Anime illustration with clean line art and cel shading.

2. A quiet scene on Spicy Imagine Pro, the Seedream 5.0 model here (2K, 8 credits) Tags: male focus, adult, short black hair, glasses, white shirt, sleeves rolled up, sitting, cafe, window, rain, warm lighting, from side, depth of field Sentences: A man in his thirties with short black hair and round glasses sits at a café window, his white shirtsleeves rolled to the elbow and a cup of coffee between his hands. Seen from the side, with rain streaking the glass and warm lamp light inside. Painterly anime style, the street behind him softly out of focus.

3. A costume change on Wan 2.7 Image (6 credits) Attach your best portrait from example 1 as the reference image. Wan 2.7 Image sits in the Image Lab’s Spicy mode, which asks for a one-time age confirmation. Tags: same character, red dress, evening gown, full body, standing, ballroom, chandelier Sentences: Keep the woman from the reference image, with the same face, black hair and amber eyes, and dress her in a long red evening gown. Show her full length, standing in a ballroom under a crystal chandelier. Keep the same anime style and line work.

Together the three cost 26 credits. To test wording, draft on Spicy Imagine Pro first (8 credits at 2K, less than the 12 of Nano Banana 2.1), then render the keeper on the model whose look you prefer.

Once a character looks right, stop rewriting the description and start attaching pictures of it. The reference slots: 14 on Nano Banana 2.1, 10 on Spicy Imagine Pro, 9 on Wan 2.7 Image. With several attached, say which picture you mean; on Spicy Imagine Pro, number them Figure 1, Figure 2, Figure 3 in upload order.

Change one thing per pass. “The same character as Figure 1, now in a grey raincoat” holds the design far better than a full description with one word changed. For new angles of a finished character, Multi-Shot makes several views from a single image.

A detail is ignored. Move it into the first sentence, or give it a sentence of its own.

Two characters swap traits. Describe each one in a separate sentence and place them: the woman on the left…, the man on the right….

The style drifts towards a photo. Name the medium early (a 2D anime illustration, cel shading, clean line art) and drop photographic words such as 85mm lens or bokeh.

Text or a watermark appears. Ask for what you want instead: a clean background with no lettering. Google’s best-practice notes for its Gemini image models, the family behind Nano Banana, give the same advice: describe what you want, not what you don’t.

The prompt has grown past a paragraph. Cut adjectives that repeat each other, and keep only what you would notice at a glance.