Anima Prompting

Anima Prompting

Anima is a 2B anime / illustration text-to-image model (CircleStone Labs + Comfy Org), built on NVIDIA Cosmos with a qwen-text-encoder LLM adapter that has outsized influence on the generated image. Write the prompt and negative you pass to the generate_image tool per the rules below to drive Anima well on the first call. The score tags, @artist rule, and qwen-text-encoder behavior below are specific to Anima.

Anima is trained on Danbooru-style tags of the Gelbooru flavor, natural-language captions, and mixes of the two (anime training data cuts off September 2025). All three work; pick whichever fits the request and always lead with quality + safety tags. Because of the qwen LLM text-encoder adapter, natural language is a first-class input mode, not a fallback - prose (or a quality-tag prefix followed by prose) often captures a scene's composition, spatial layout, and mood better than tags alone. The adapter is what makes detailed prompts pay off here far more than on a plain SDXL model - terse prompts waste the model's strongest feature.

Safe default when unsure: use the anima workflow, begin the prompt with masterpiece, best quality, score_7, safe,, describe the subject in roughly 12-25 lowercase tags (or 2+ sentences of prose), and always pass the recommended negative from the negative recipe. Short prompts make Anima bland or unsafe, so spend the detail. Everything below is how to do better than this default when the request calls for it.

The generate_image call

You set the entire prompt string - it replaces the workflow's baked text, so the quality/safety prefix must be in your prompt every time. Do not pass inputImage (Anima here is text-to-image only).

Positive prompt recipe

1. Lead with the prefix: masterpiece, best quality, score_7, safe,
2. Then the subject, as tags, natural language, or a mix.
3. Lowercase tags, spaces not underscores - the only underscored tags are score tags (score_7). Prefer the Gelbooru spelling when a tag differs between boorus.
4. Be specific. Anima's base style is plain; short prompts give bland or unexpected (sometimes unsafe) output. For natural language aim for >=2 sentences. For tag style aim for >=8 substantive tags after the prefix. The model card warns explicitly: "The model may generate undesired content, especially if the prompt is short or lacking details."
5. Name then describe characters. Fern from Sousou no Frieren, with long purple hair and purple eyes, wearing a black coat... - especially with multiple characters, describe each one's appearance or the model conflates them.
6. Artists need @: write @artist name (the @ is mandatory; without it the style barely registers). You may put quality/artist tags at the start of a natural-language prompt too. An @artist tag also locks a consistent style across seeds - pin one when you want the same house look from image to image (verified: same @artist holds the style steady across different seeds).

Tag order (tag-style prompts)

[quality / meta / year / safety] [1girl / 1boy / 1other] [character] [series] [@artist] [general tags]

Order only matters between sections; within a section tags are free-order. You do not need every relevant tag - the model was trained with tag dropout.

Natural-language and mixed prompts

Anima's qwen text-encoder adapter reads full sentences, not just tags, so natural language is a first-class way to prompt it - often the better one for scenes, spatial relationships, mood, and multi-element compositions where a tag list gets ambiguous. The model card ships a dedicated natural-language tips section; the rules that matter:

• Keep the quality / safety prefix. Lead with masterpiece, best quality, score_7, safe, (or .) even in a prose prompt - the rating tag still carries safety - then write the description as prose. Quality and @artist tags are fine at the start of a natural-language prompt.
• Be descriptive: aim for >=2 sentences. Pure natural language rewards detail; extremely short prose gives unexpected (sometimes unsafe) results. Spend the same detail budget as the tag recipe (~12-25 depictable items), just expressed as prose - appearance, wardrobe, setting, lighting, palette.
• Mix tags and prose freely, in any order. A hybrid is common and effective: open with quality + character / series tags, then a sentence or two of prose for pose, framing, and scene. The tool takes one string; comma- or period-join the two halves.
• Name a character, then describe their appearance (Fern from Sousou no Frieren, with long purple hair and purple eyes, wearing a black coat...). With multiple characters this is essential or the model conflates them.

The one caution: do not bury a single character under a wall of scenery prose - keep scene description to a short clause and spend the detail on the character (see Anti-patterns).

Expanding a terse user request

When the user asks for something short like "draw a cat girl" or "make me a witch", do not pass that string through. Expand it before calling generate_image:

1. Prefix: add masterpiece, best quality, score_7, safe,.
2. Subject count + type: infer 1girl / 1boy / 1other, plus solo if a single character.
3. Character / series: only if the user named one (Frieren from Sousou no Frieren); otherwise skip.
4. Appearance: invent 3-5 concrete details - hair color + length, eye color, outfit, distinctive accessory. Match the user's vibe; ask only if their wording is genuinely ambiguous about a load-bearing detail (e.g. "anime girl in a kimono - red or blue?").
5. Pose / framing: 1-2 tags - looking at viewer, standing, upper body, cowboy shot. Pick ONE framing tag.
6. Setting: 2-4 tags or a clause - sunlit library, tall window, dust motes, or simple background, white background for a clean character shot.
7. Lighting + palette: 1-3 tags - soft lighting, cinematic lighting, warm palette, muted colors.
8. Era (optional): add year 2025 or newest for a modern look; skip for timeless.
9. Artist (optional): only if the user named one - prepend @ (e.g. @some artist).
10. Negative: always pass the recommended negative; add extra fingers, text, watermark for cleaner output.

The goal is roughly 12-25 substantive items in the positive prompt. If the user wants to iterate, pass the prior seed back to vary by tweaking tags rather than restarting.

When the user supplies tags directly

If the user hands you a tag list (or a tag list mixed with wildcards) instead of a terse English request, switch modes:

1. Preserve every user-supplied tag verbatim. Do not paraphrase, deduplicate, reorder across sections, or "improve" their wording.
2. Prepend the prefix if missing. Add masterpiece, best quality, score_7, safe, only if the user did not already include quality / safety tags.
3. Add spatial / positional words only. Insert framing words like in the center of the frame, foreground, background, in front of, surrounded by, on either side to disambiguate where elements sit. Do not add weather, lighting, palette, or outfit details the user did not ask for.
4. Preserve dynamic-prompt wildcards exactly. {A|B}, {A,|B,|C}, {1-3$$ A|B|C}, and {A,B}_noun are wildcard-extension syntax (ComfyUI Impact-Pack, sd-dynamic-prompts) - they expand at generation time, not in your prompt. Pass them through unchanged: never rewrite {standing|sitting} as "standing or sitting", never pick one branch, never delete a branch.
5. Append a one-sentence natural-language clause describing spatial relationships - Anima's qwen adapter benefits from prose, so a single closing sentence on top of the tag list usually improves coherence without adding embellishment.

prompt:   masterpiece, best quality, score_7, safe, 1girl, {standing|sitting}, classroom, desk, {morning|evening},
          a young female student positioned in the center of the classroom in front of the desk, with the
          {morning|evening} lighting implied by the scene
negative: worst quality, low quality, score_1, score_2, score_3, artist name

Concatenate the tag list and the natural-language clause into one prompt string (comma- or period-joined) - the tool takes a single string, and Anima reads mixed input fine.

Negative prompt recipe

Start from the official recommendation and add situational terms:

worst quality, low quality, score_1, score_2, score_3, artist name

• Add jpeg artifacts, blurry, lowres to clean up artifacts.
• Add the specific thing you don't want (e.g. multiple views, text, watermark, extra fingers).
• Do not put a rating word (safe, sensitive, nsfw, explicit) in the negative. Safety is carried by the rating tag in the positive prefix, not by negating it here, and putting nsfw in the negative actively fights the positive the moment the scene is sensitive or higher. Keep the negative rating-free; every example below does. (On a safe shot, on a workflow that does not auto-scrub the negative, you may reinforce with nsfw here - but never above safe. When in doubt, leave it out and let the positive rating tag do the work.)

Tag reference

Weighting syntax. Anima runs on ComfyUI, which weights with parentheses only: (tag:1.4) to strengthen, (tag:0.6) to weaken, bare (tag) for ~1.1x, and nesting (((tag)) ~= 1.21x). Anima responds more weakly to a given weight than SDXL, so reach for bigger moves - (chibi:2), not (chibi:1.2). Square brackets are not weighting on ComfyUI: [tag] is parsed as ([tag]:1), i.e. the brackets become literal tokens at weight 1, not an A1111-style de-emphasis. To push something down, use a fractional parenthesis weight like (background:0.6), never [background].

Generation settings

Sampler/scheduler live in the workflow file, not the tool args - mention a sampler only if asking the user to retune the workflow.

Prompt enhancement (enhance)

The generate_image tool has an opt-in enhance option that routes your prompt through a separate model before the render to rewrite it into Anima's native protocol - the score/safety-prefixed tag list this guide describes.

• It is opt-in. Pass enhance: true on the call (or a workflow/config default turns it on). When off, your prompt is sent as-is.
• Use it when the incoming prompt is thin (a few words) and you want a one-shot upgrade to a full tagged prompt. When you have already built a careful tagged prompt, leave enhance off - it only adds latency and risks drifting from your intent.
• Scene continuity. Pass the context arg alongside enhance to hand the enhancer background to honour (character facts, ongoing scene, wardrobe) without depicting it literally. It is ignored when enhance is off.
• Negative. The enhancer builds on whatever baseline negative you pass and returns a refined one - it should keep rating words out of the negative (see the negative recipe).

If you are enhancing a prompt rather than rendering

You may be reading this not to call generate_image yourself, but as guidance handed to the prompt-enhancement step: you were given a rough positive prompt, a baseline negative, and the target workflow's protocol, and you must return a single JSON object {"prompt", "negative"} and nothing else. In that role:

• Act only on the prompt-writing rules above - the score/safety prefix, the positive recipe and tag order, the terse-request expansion, and the negative recipe. The generation-settings and tool-arg tables do not apply - you do not pick a workflow or call any tool.
• Translate and enrich, do not reinvent. Keep the incoming subject and intent; lead with masterpiece, best quality, score_7, safe, and turn loose phrasing into lowercase Gelbooru-style tags (or a tags-plus-prose mix). Preserve any dynamic-prompt wildcards ({A|B}) unchanged.
• Always add detail - even to an already-tagged prompt. Anima produces bland or unsafe output from short prompts, so push toward roughly 12-25 substantive items: infer subject count (1girl / 1boy / 1other, solo), then layer in appearance, one framing tag, setting, and lighting / palette that the prompt leaves unstated, staying faithful to the stated subject.
• Mine any background context you were given (a scene/continuity note, recent conversation) for that extra detail and fold it in as depictable tags. Treat it as source material to pick from, not a checklist to dump and not a subject that overrides the explicit prompt; ignore chatter with no visual bearing.
• Build on the baseline negative (start from worst quality, low quality, score_1, score_2, score_3, artist name), and keep rating words (safe / sensitive / nsfw / explicit) out of it.

Worked examples

These illustrate the shape of a finished prompt - adapt them to the actual request, do not reuse them verbatim.

Tag-style:

prompt:   masterpiece, best quality, score_7, safe, 1girl, solo, long hair, brown eyes, santa costume,
          fur-trimmed gloves, holding gift box, looking at viewer, simple background, white background, @some artist
negative: worst quality, low quality, score_1, score_2, score_3, artist name, jpeg artifacts

Natural-language:

prompt:   masterpiece, best quality, score_7, safe. A young anime witch with short silver hair and green eyes
          stands in a sunlit library, wearing a wide-brimmed black hat and a deep blue robe. She holds an open
          spellbook that glows faintly, and dust motes drift in the light from a tall window behind her.
negative: worst quality, low quality, score_1, score_2, score_3, artist name, blurry

The wrapped lines above are a single comma/space-joined string each - pass them as one line in the tool args.

Anti-patterns

• Asking for realism / photos. Anima is anime/illustration only and will not do realism - redirect or set expectations instead of fighting it.

• Heavy environment prose on a single-character shot. A long multi-sentence paragraph describing the room/scene dilutes the character tags and visibly degrades character detail. Keep prose to a short clause (one sentence) for pose / light / spatial framing and describe the character with tags. Character-focused prose is fine; a wall of scenery prose is not. (Verified by A/B at a fixed seed.)

• Dropping the setting entirely. If neither a tag nor the prose names the location, Anima defaults to a plain white background. Name the setting somewhere if you want one.

• Three-word prompts. a cat girl underuses the model and risks plain or unsafe output. Run it through the expansion algorithm - target ~12-25 substantive items. As a sketch, a cat girl becomes:

  masterpiece, best quality, score_7, safe, 1girl, solo, cat ears, cat tail, short brown hair, green eyes,
  casual hoodie, sitting cross-legged, looking at viewer, simple background, soft lighting, newest
  

• Underscores in tags. brown_hair is wrong; write brown hair. Only score_7 keeps the underscore.

• SDXL-level weights. (chibi:1.2) barely moves Anima; use (chibi:2). And weight only with parentheses - [chibi] does not de-emphasize on ComfyUI (it is read as ([chibi]:1)); use (chibi:0.6) to weaken instead.

• Forgetting @ on an artist. some artist ~= no effect; @some artist applies the style.

• Expanding wildcards. {standing|sitting} is dynamic-prompt syntax that resolves at generation time. Rewriting it as "standing or sitting", picking one branch, or deleting it are all wrong - pass the literal {...|...} token through unchanged.

• Paraphrasing user-supplied tags. When the user gives you a tag list, do not "improve" it - preserve their tags verbatim and only add spatial framing words. See When the user supplies tags directly.

• Long text in the image. Anima renders a single word or short phrase at best - do not promise a paragraph on a sign.

• Dropping the negative arg. Always pass one; it is the main lever for quality and safety.