Quick & Nasty Prompt Design Guide

Illustrious + Anima

The golden rule: Prompt in blocks, not soup.
Organize related concepts together, separate major ideas cleanly, and only add complexity when the image actually needs it.

⚡ 30-Second Cheat Sheet

                          **Illustrious**         **Anima**

  Best starting language  Danbooru tags           Tags + natural language

  Character traits        Tags                    Tags

  Complex actions         Tags first              Natural language often
                                                  helps

  Relationships / spatial Can be finicky          Describe them naturally
  instructions                                    

  Long prompts            Split into logical      Split into logical
                          blocks                  blocks

  Core philosophy         **TAG IT.**             **TAG IT, THEN EXPLAIN
                                                  THE WEIRD PART.**

Recommended Prompt Order

Quality → Subject → Appearance → Clothing → Pose / Expression → Scene → Camera → Lighting / Style

Keep each character's important traits together. Don't scatter hair, eyes, clothes, and body traits across the entire prompt.

🎨 Illustrious

Illustrious generally responds extremely well to Danbooru-style tags. Short, known concepts are its native tongue.

Basic Structure

masterpiece, best_quality, amazing_quality, very_aesthetic, absurdres,
1girl, solo, adult_woman, long_black_hair, blue_eyes, athletic_build,

BREAK

black_dress, thighhighs, jewelry,
smile, looking_at_viewer, standing, hand_on_hip,

BREAK

night, city_street, neon_lights, rain,
cinematic_lighting, dramatic_shadows, depth_of_field

Rules

1. Tags Beat Prose

Prefer known tags:

long_silver_hair, red_eyes, black_dress

instead of unnecessarily describing a simple concept in prose.

Save natural language for something tags cannot communicate cleanly.

2. Important Concepts Go Early

The front of the prompt is premium real estate.

Prioritize:

1. Subject
2. Defining appearance
3. Critical clothing / objects
4. Pose and expression
5. Environment
6. Decorative details

3. Use BREAK to Divide Concepts

A very useful practical rule:

Consider a BREAK roughly every 75 tokens when working with CLIP-style text-encoder blocks.

Treat 75 as a working guideline, not sacred mathematics. The point is to keep a huge prompt from becoming one conceptual casserole.

A clean pattern:

quality + character

BREAK

clothing + pose + expression

BREAK

environment + composition

BREAK

lighting + artistic treatment

4. Weight Gently

If the model repeatedly ignores an important trait:

(red_eyes:1.1)
(red_eyes:1.15)
(red_eyes:1.2)

Increase gradually.

If half your prompt is weighted, the weights stop being useful. Use them as a scalpel, not a sledgehammer.

5. Don't Feed It Contradictions

Bad:

smile, serious_expression, angry, happy

Decide what you actually want.

6. Keep Negatives Useful

Start relatively lean:

worst_quality, low_quality, bad_anatomy, bad_hands,
extra_fingers, missing_fingers, text, watermark

Then add negatives to solve recurring problems rather than pasting an enormous generic negative prompt into every generation.

🏷️ Danbooru Tags: Use the Actual Tag

For multi-word Danbooru tags, prefer the tag's real underscore formatting.

❌ long silver hair
✅ long_silver_hair

❌ looking at viewer
✅ looking_at_viewer

❌ cowboy shot
✅ cowboy_shot

❌ depth of field
✅ depth_of_field

Danbooru tags are named concepts rather than arbitrary English phrases. When you know the actual tag, use it.

But Don't Invent Tags

This:

woman_standing_in_a_beautiful_room_holding_a_glass

does not become a useful Danbooru tag merely because you attacked the spacebar with underscores. 😈

Use underscores for real tags.

For complicated ideas, use natural language where the model supports it:

a woman standing beside the window, holding a glass in one hand

Easy Rule

Known Danbooru concept? → use_the_tag

Complex relationship or action? → describe it naturally when appropriate

Illustrious? → favor tags heavily

Anima? → tags for anchors, language for complicated relationships

🌸 Anima

Anima can benefit from a hybrid approach: strong tags establish the concepts, while natural language explains relationships, actions, and spatial details.

Basic Structure

masterpiece, best_quality, very_aesthetic, 1girl, solo,
long_silver_hair, red_eyes,

a confident adult woman with elegant facial features,

BREAK

wearing a fitted black evening dress with subtle silver jewelry,
standing beside a rain-covered window,
looking toward the viewer with a faint smile,

BREAK

luxurious hotel interior at night,
city lights visible outside,

BREAK

cinematic lighting, soft rim light,
detailed background, shallow depth of field

Rules

1. Tags Establish Anchors

Use tags for concepts the model already understands:

1girl, solo, silver_hair, red_eyes, black_dress

2. Language Establishes Relationships

Natural language becomes valuable when how things relate matters:

her hair blowing across her face

holding the umbrella above both characters

looking over her shoulder toward the viewer

one hand resting against the window

This can communicate more clearly than piling on another twenty isolated tags.

3. Use BREAK for Conceptual Separation

The same practical block structure works beautifully:

quality + character

BREAK

clothing + pose + expression

BREAK

environment + interaction

BREAK

camera + lighting + atmosphere

4. Don't Over-Tag

More tags do not automatically equal more control.

If the prompt already establishes:

woman, dress, room, window, rain

and you need a very particular interaction, describe the interaction:

the woman stands directly beside the window,
touching the rain-covered glass with her fingertips

Specific relationships often beat tag avalanches.

🧱 Character Consistency

Keep permanent character traits in one logical block.

Good:

1girl, long_pink_hair, blue_eyes, freckles, curvy

Messier:

1girl, pink_hair, city, dress, blue_eyes, rain, freckles...

The cleaner the character definition, the easier it is to diagnose what went wrong.

For multiple characters, keep each character's traits as grouped as your workflow/model allows. Attribute bleeding is the tiny goblin hiding behind many multi-character failures.

📷 Composition Is Part of the Prompt

Useful composition concepts include:

close-up
portrait
upper_body
cowboy_shot
full_body
wide_shot
from_above
from_below
dutch_angle
depth_of_field
foreground
background

A surprising number of apparent character failures are really composition failures.

If the model keeps cropping shoes, don't spend twelve generations yelling about shoes. Tell it what shot you want.

🎭 Describe Actions Precisely

Weak:

woman sitting

Better:

woman sitting on sofa

Better still:

woman sitting sideways on a sofa,
legs crossed,
one arm resting along the backrest

Every useful spatial relationship removes another decision from the model.

⚖️ Weighting Cheat Sheet

(concept:1.05)  tiny nudge
(concept:1.1)   gentle emphasis
(concept:1.2)   noticeable emphasis
(concept:1.3)   strong emphasis

Treat these as rough working ranges. Different checkpoints, UIs, encoders, and workflows can respond differently.

Start low. Increase only when the model proves stubborn.

🔧 Emergency Prompt Repair

When the image is wrong, don't rewrite everything.

Problem First thing to change

Wrong hair / eyes Character block Wrong clothing Clothing block Wrong expression Expression / pose block Wrong pose Pose description Missing important object Move it earlier or gently weight it Character too small Composition / framing Background dominates Reduce background detail or move it later Attributes bleeding Separate and simplify character definitions Prompt feels incoherent Remove contradictions and unnecessary tags Long prompt losing concepts Reorganize with BREAK

Prompt debugging should be surgical.

Change one meaningful variable, generate again, and see what moved.

Do not add 47 tags and pray to the GPU. 🔧🔥

📋 Copy-Paste Templates

Illustrious

masterpiece, best_quality, very_aesthetic,
[subject], [appearance],

BREAK

[clothing], [expression], [pose],

BREAK

[environment], [composition],

BREAK

[lighting], [style]

Anima

masterpiece, best_quality, very_aesthetic,
[basic subject tags],
[short natural-language character description],

BREAK

[clothing + precise action / pose],

BREAK

[environment + relationship to environment],

BREAK

[camera + lighting + atmosphere + style]

🧪 The Actual Workflow

1. Start simple.
2. Generate.
3. Identify the specific failure.
4. Find the block responsible.
5. Change that block.
6. Generate again.
7. Add complexity only when it earns its keep.

A clean 40-tag prompt you understand is vastly easier to tune than a 250-tag eldritch incantation copied from somebody else's metadata.

🧠 Final Memory Trick

Illustrious

TAG IT. GROUP IT. BREAK IT.

Anima

TAG THE CONCEPT. DESCRIBE THE RELATIONSHIP.

And for both:

If the model gets something wrong, fix the instruction responsible instead of making the entire prompt louder.