Hello everyone! Last year, I published The Ultimate Illustrious XL Prompting Guidebook, sharing what I had learned from working with Illustrious. And I was genuinely grateful to see it receive so many likes and such positive feedback from the community. That support encouraged me to return this year with a new guide for Anima, a newer anime model that has now been around for a few months.

During that time, I have studied real prompts, compared results, experimented with different prompting styles, and followed what other users have discovered. Anima can understand both tags and natural language, but it does not always behave like Illustrious. Getting good results often depends not only on what you describe, but also on how clearly you communicate the composition, lighting, color, materials, and overall visual idea.

This guide collects everything I have learned so far. It covers Anima’s prompting basics, useful community techniques, visual concept development, practical examples, and common mistakes. The model is still evolving, so this is not meant to be the final answer. It is simply the most useful and honest summary I can offer based on what we currently know.
If you enjoy my knowledge-sharing articles and would like to support my work, you can visit my Tensor.Art profile and purchase a few of my models. Your support would mean a lot to me and give me even more motivation to continue writing guides about new models such as Krea.

Anima Prompting Guidebook

A practical, evidence-based handbook for official Anima 2B models

Coverage: Anima-Base v1.0, Anima-Aesthetic v1.1, and Anima-Turbo v1.1

Primary focus: prompt structure, model settings, style control, multi-character scenes, testing, and troubleshooting

Research standard: official documentation is separated from community-tested practice

Table of Contents

The guide moves from model choice and syntax to controlled testing, migration, derivatives, and complete prompt examples.

1. What Anima Actually Is

2. Choosing Base, Aesthetic, or Turbo

3. Installation and Baseline Generation Settings

4. The Three Prompting Modes

5. The Anima Hybrid Prompt Formula

6. Tag Syntax, Order, and Formatting

7. Quality, Year, Meta, and Safety Tags

8. Character and Original-Character Prompting

9. Artist Tags and Style Control

10. Composition, Camera, Lighting, and Backgrounds

11. Multi-Character Prompting

12. Prompt Weighting and Negative Prompts

13. Testing and Troubleshooting

14. A Small Illustrious-to-Anima Migration Guide

15. Notable Anima Derivatives

16. Complete Prompt Examples

17. Limitations and Final Checklist

1. What Anima Actually Is

Anima is an anime- and illustration-focused text-to-image model created by CircleStone Labs in collaboration with Comfy Org. It has a 2B diffusion model and uses a Qwen-3 0.6B text encoder in the official ComfyUI implementation. The diffusion backbone is derived from NVIDIA's Cosmos-Predict2-2B-Text2Image rather than Stable Diffusion XL. This distinction matters: Anima is not an SDXL finetune wearing a new name, and SDXL or Illustrious LoRAs should not be assumed to work with it.[1][2][3]

The official model card states that Anima was trained on several million anime images and roughly 800,000 non-anime artistic images, without synthetic training data. Its anime-data knowledge cutoff is September 2025. It is intended for anime, illustration, concept art, and other non-photorealistic work; realism is explicitly outside the official model's main target.[1]

The official weights and derivative models use the CircleStone Labs Non-Commercial License, while the model card states that generated outputs may be used commercially. Read the license itself before distributing weights, offering hosted inference, or embedding the model in a product.[1]

Anima learned three kinds of text conditioning:

• Danbooru-style visual tags

• Natural-language captions

• Mixtures of tags and natural-language captions

That third category is the most important practical difference in how the model should be approached. Tags remain useful for exact visual attributes, while prose is useful for relationships, staging, and scene logic.[1]

What the Qwen text encoder changes in practice

The text encoder gives Anima stronger sentence-level interpretation than creators may expect from older tag-first anime checkpoints. This does not mean that prose automatically beats tags. It means that each format has a different job:

• Use tags to anchor count, identity, visible traits, clothing, expression, medium, and known visual concepts.

• Use natural language to explain who is doing what, where each subject is located, how objects relate, and what the scene should communicate.

• Use both when the scene has precise appearance requirements and non-trivial spatial logic.

The official documentation also says that pure natural-language prompts should normally contain at least two descriptive sentences. Extremely short prose prompts can produce unexpected content or an under-specified composition.[1]

2. Choosing Base, Aesthetic, or Turbo

The three official versions share the same Anima prompt language, but they do not have the same default behavior.

Anima-Base

v1.0

Maximum diversity, artist adherence, LoRA training, neutral testing

Plain default style; needs deliberate style or quality guidance

Anima-Aesthetic

v1.1

Polished single images and a stronger default anime finish

Less neutral than Base; score tags can over-push the aesthetic

Anima-Turbo

v1.1

Fast prompt iteration and inexpensive seed exploration

Stronger default style and less diversity than Base

Official: Base is the unrefined foundation and the recommended target for LoRA training. Aesthetic is tuned for consistency and a higher-quality default. Turbo is distilled for speed and is officially run at CFG 1 for 8–12 steps. CircleStone describes Turbo as more stable but less diverse, and recommends it as a convenient first version for rapid iteration.[1]

The latest Civitai release notes add two useful details. Aesthetic v1.1 was partially retrained for smoother styling and fewer artifacts. Turbo v1.1 retrained much of the distillation and style-adjustment stack to produce less-flat results, more detail, and somewhat better artist-style response.[4]

Practical selection

• Choose Base when you want to discover styles, compare artist tags, train a LoRA, or keep the model as unbiased as possible.

• Choose Aesthetic when you want a dependable polished illustration without building the entire aesthetic from scratch.

• Choose Turbo when you want to test many prompt revisions or seeds quickly.

Do not evaluate prompt syntax while changing versions, samplers, CFG, resolution, and seed at the same time. First establish a working baseline on one version.

3. Installation and Baseline Generation Settings

Anima is natively supported by ComfyUI. The official Base workflow uses three separate files:

• anima-base-v1.0.safetensors in ComfyUI/models/diffusion_models

• qwen_3_06b_base.safetensors in ComfyUI/models/text_encoders

• qwen_image_vae.safetensors in ComfyUI/models/vae

The official ComfyUI Template Library contains an “Anima Base v1” workflow. Starting from that template is safer than converting an old SDXL graph node by node.[2]

Official generation range

The current official model information gives the following baseline:

Resolution

Between 512×512 and 1536×1536

Start within the official model range

Steps

30–50

8–12

CFG

4–6 on the current Civitai card; the Hugging Face card lists 4–5

1

Safe default sampler

er_sde

euler or another sampler tested for Turbo

The small CFG discrepancy comes from two official pages. A good neutral starting point for Base is therefore CFG 4.5, not an imported SDXL value such as 7 or 8.[1][4]

Official sampler characteristics

• er_sde: neutral style, flat color, and sharp linework; the developer's default recommendation.

• euler_a: softer, thinner lines and sometimes a more 2.5D appearance; it can tolerate somewhat higher CFG.

• dpmpp_2m_sde_gpu: similar to er_sde but more varied and creative; it may become too wild on some prompts.

• euler: somewhat more creative than er_sde and a useful option for the naturally stable Aesthetic and Turbo versions.

• beta57 scheduler: an optional choice for painterly or semi-realistic texture because it emphasizes lower-noise timesteps. It is not required for ordinary anime illustration.[1]

A safe first test

Use this before adding LoRAs, upscalers, detailers, or experimental schedulers:

Parameter

Baseline

Version

Anima-Base v1.0

Resolution

1024×1024

Steps

35

CFG

4.5

Sampler

er_sde

Scheduler

Use the official workflow default

Seed

Fixed while editing the prompt

If this baseline fails completely, first verify the text encoder and VAE. A wrong architecture component can make prompt work look like a prompting problem.

4. The Three Prompting Modes

Mode A — Tag-only

Tag-only prompts are compact and precise. They work well for a single subject, familiar poses, known clothing, and simple compositions.

masterpiece, best quality, score_7, safe, year 2025, highres,
1girl, solo, silver hair, short hair, blue eyes, black sailor uniform,
standing, looking at viewer, gentle smile, train platform, sunset,
backlighting, wind, anime coloring

Strengths:

• Fast to write and edit

• Easy to A/B test

• Strong control over canonical visual attributes

Weaknesses:

• Relationships can remain ambiguous

• Long tag chains can create contradictions

• Spatial instructions such as “behind,” “to the left,” or “holding one object with both hands” may be clearer in prose

Mode B — Natural language

Pure prose is useful for scenes whose meaning depends on relationships or action.

A young anime swordswoman stands alone on an abandoned railway platform at sunset. She holds a closed umbrella in her right hand while the wind blows her short silver hair toward the left, and warm backlight outlines her dark sailor uniform. Sword on her back.

Use ordinary English capitalization for character and series names. For pure natural language, write at least two concrete sentences and describe visible information rather than lore or personality that has no visual consequence.[1]

Mode C — Hybrid

Hybrid prompting is usually the most reliable general-purpose method:

masterpiece, best quality, safe, year 2025, highres, 1girl, solo,
silver hair, short hair, blue eyes, black sailor uniform, anime coloring.
A young swordswoman stands at the center of an abandoned railway platform at sunset. She holds a closed umbrella in her right hand while the wind carries her hair and loose ribbons toward the left; warm rim light separates her silhouette from the tracks behind her.

The tag block establishes the visual vocabulary. The prose block assigns those elements to a composition.

Community-tested: creators repeatedly report that separating “what must be present” from “where it goes and what it is doing” makes complex prompts easier to debug. This is consistent with the official hybrid-training description, but the exact line breaks are an organizational convention, not a required hidden syntax.[5][6]

5. The Anima Hybrid Prompt Formula

The official tag order is:

[quality / meta / year / safety] → [subject count] → [character] → [series] → [artist] → [general tags][1]

For practical hybrid prompting, extend it with a natural-language scene block:

Quality + era + safety → subject count → identity → series → artist/style → visible attributes → pose/action → camera → environment → lighting → natural-language spatial description

Reusable template

[quality tags], [year or period], [meta tags], [safety tag],
[subject count],
[character name], [series],
[@artist tag or medium/style],
[hair], [eyes], [clothing], [expression],
[pose/action], [camera/framing],
[environment], [lighting/effects].

[Sentence 1: identify the main subject and place them in the frame.]
[Sentence 2: describe the action, object relationships, depth, and important lighting.]

Minimal hybrid template for an original character

safe, highres, 1girl, solo, red hair, braided hair, green eyes,
white mage robe, leather satchel, full body, forest path, morning light.

An apprentice mage walks along the center of a mossy forest path while holding an open map with both hands. Tall trees recede behind her, and narrow shafts of morning light fall through the canopy onto the path.

This formula is deliberately modular. If a result changes in an unwanted way, remove or replace one module instead of rewriting the entire prompt.

6. Tag Syntax, Order, and Formatting

Use lowercase tags

Write visual tags in lowercase:

• Correct: blue hair, long sleeves, looking at viewer

• Avoid: Blue Hair, Long Sleeves, Looking At Viewer

Natural-language sentences and proper nouns are different: normal English capitalization is appropriate there.[1]

Use spaces, not underscores

Anima's official rule is to use spaces in ordinary tags. Score tags are the exception.

• Correct: blue hair, upper body, from above, score_7

• Avoid: blue_hair, upper_body, from_above

When Danbooru and Gelbooru use different tag forms, the official model card recommends the Gelbooru form.[1]

ASCII emoticons and Booru emote tags

Anima can interpret common ASCII-style strings when they are canonical Booru expression tags; this does not imply support for arbitrary ASCII art. The official full-tag example uses ;d, while independent corpus tests include Anima generations for :3 and >_<.[1][17][18]

safe, highres, 1girl, solo, chibi, :3, blush, cat mouth,
looking at viewer, simple background, anime coloring

safe, highres, 1girl, solo, >_<, closed eyes, open mouth,
laughing, hands on cheeks, upper body, colorful background

Keep the canonical spelling and punctuation. The underscore in >_< is part of the tag name, so do not normalize it to spaces. Use one primary emote and pair it with plain expression tags such as closed eyes, laughing, or cat mouth. Stacking incompatible emotes makes the intended eyes and mouth ambiguous.

These tags control the character's visual expression; they do not reliably print the emoticon as literal text inside the image. Anima may render a single word or short phrase, but the official model card warns that longer text is unreliable.[1]

Put spaces after commas

Write:

1girl, solo, blue hair, white dress

Do not compress the list into:

1girl,solo,blue hair,white dress

The official card does not describe comma spacing as a special control mechanism, but clear separation prevents accidental token merging and is consistently recommended by experienced users.[5]

Do not tag every invisible fact

Anima was trained with random tag dropout. You do not need to reconstruct every tag that might have described a training image.[1]

Prioritize:

1. Subject count

2. Identity and series

3. Defining appearance

4. Action and pose

5. Camera and framing

6. Environment

7. Lighting and style

Low-value metadata and duplicated synonyms consume attention while making failures harder to diagnose.

Avoid contradictions

Do not combine incompatible instructions unless the contradiction itself is intentional:

• short hair with very long hair

• front view with from behind

• closed mouth with shouting

• solo with 2girls

Community-tested: some users find that Anima tries harder than Illustrious-style checkpoints to satisfy conflicting concepts, producing anatomical or compositional failures rather than quietly ignoring one tag. Treat this as an anecdotal warning, not a measured benchmark.[5]

7. Quality, Year, Meta, and Safety Tags

Base: start with the official prefix

The official Base recommendation is:

masterpiece, best quality, score_7, safe,

The official negative baseline is:

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, chromatic aberration

This is a starting point, not a law.[1]

Aesthetic: remove score tags

Anima-Aesthetic was fine-tuned on high-quality images with quality labels stripped from the captions. The developer says quality tags are not necessary and specifically recommends avoiding score_* tags in both positive and negative prompts. masterpiece, best quality is safe to leave in, but may be redundant.[1]

A clean Aesthetic baseline is:

safe, highres,

or:

masterpiece, best quality, safe, highres,

Why users disagree about score_7

Community reports conflict:

• Some users find score_7 necessary for a reliable high-quality Base result.

• Others find that it narrows variation and pulls the result toward a flatter Pony-like aesthetic.

• The official documentation resolves part of this disagreement by treating Base and Aesthetic differently.[5]

Use a fixed seed and compare:

1. No quality tag

2. masterpiece, best quality

3. masterpiece, best quality, score_7

Judge linework, color, anatomy, diversity, and whether the intended artist/style remains recognizable.

Year and period tags

Anima understands:

• Specific years: year 2025, year 2024

• Broad periods: newest, recent, mid, early, old

These are aesthetic controls, not factual timestamps. A year tag can shift the visual distribution, so change it separately from artist and medium tags.[1]

Useful meta tags

Official examples include:

• highres

• absurdres

• anime screenshot

• official art

• jpeg artifacts

Do not assume every impressive-sounding token improves quality. Use tags that occur in the documented vocabulary and remove any tag whose effect you have not tested.

Safety tags

The documented safety vocabulary is:

• safe

• sensitive

• nsfw

• explicit

The official card warns that short or under-specified prompts may generate unwanted content. Use the intended safety tag in the positive prompt and keep the scene sufficiently descriptive.[1]

8. Character and Original-Character Prompting

Known characters need three anchors

For a known character, include:

1. Character name

2. Series or franchise

3. Basic visible appearance

Do not rely only on the name.

safe, highres, 1girl, solo, fern, sousou no frieren,
long purple hair, purple eyes, black coat, white dress, long sleeves.

Fern from Sousou no Frieren stands in a quiet flower field. Her long purple hair falls behind her black coat, and she holds a wooden staff vertically beside her while looking toward the viewer.

The official guide explicitly recommends naming the character and describing their basic appearance. This becomes more important with multiple characters, where a bare list of names can confuse identity assignment.[1]

Use canonical spelling

For the tag block, use the canonical lowercase tag or the exact entry from an Anima-specific autocomplete list. For prose, use standard capitalization. Community-maintained Anima 1.0 lists use Gelbooru-based tags, spaces, a September 2025 cutoff, and the required @ artist prefix. They are useful autocomplete references, but they are not proof that every listed concept is equally strong in the model.[7]

Original characters

Original characters have no built-in identity anchor. Define them with a compact “identity packet”:

short auburn hair, single braid, amber eyes, round glasses,
navy school blazer, red ribbon, pleated skirt, star-shaped hairpin

Keep this packet stable across generations. Put temporary pose, weather, and camera instructions outside it.

Bad:

auburn hair, red hair, orange hair, short hair, waist-length hair,
glasses, no glasses, school uniform, fantasy armor

Better:

short auburn hair, single side braid, amber eyes, round glasses,
navy school blazer, red ribbon, pleated skirt, star-shaped hairpin

Consistency is not guaranteed by repeated text

A text prompt describes a concept; it does not create a persistent character identity. If exact face and costume consistency matter across many images, use an Anima-trained character LoRA or an appropriate reference workflow. Do not import an Illustrious character LoRA and assume architectural compatibility.

9. Artist Tags and Style Control

The @ prefix is mandatory

Artist tags must begin with @:

• Correct: @kantoku

• Weak or incorrect: kantoku

The official card says the influence becomes very weak without the prefix.[1]

Use canonical lowercase spelling. A code-level tokenizer inspection found that differently capitalized artist strings produce different token sequences; this supports the practical rule to copy the canonical tag exactly.[8]

Start with one artist

The most reliable style-discovery workflow is:

1. Fix prompt, seed, model version, sampler, and settings.

2. Test without an artist tag.

3. Add one verified @artist tag.

4. Generate a seed grid.

5. Only then test a second artist or a style LoRA.

This separates the artist's actual influence from the model's default seed variation.

Artist tags are not universal style presets

An artist tag reflects the distribution and labeling of that artist's training examples. Some tags are strong, some weak, and some carry content biases such as a repeated camera angle, expression, costume, or rendering medium.

Community-tested: users report style drift across seeds even with one artist, and larger drift when several artists are mixed. A character always looking away, a repeated costume, or unexpected realism may come from an artist tag or an overfit LoRA rather than the subject prompt.[9]

Weighting artist tags

The official model supports weights and notes that Anima often needs stronger values than SDXL. Test conservatively:

(@kantoku:1.5)

then:

(@kantoku:2)

Do not jump directly to extreme values or weight five artists at once. Strong weights can alter composition and anatomy, not only rendering style.

Style tags can be more stable than artist cocktails

For reproducible art direction, define the visual target directly:

anime coloring, clean lineart, flat colors, soft cel shading,
muted pastel palette, thin outlines, official art

Then use one artist tag as an accent rather than asking several artist identities to negotiate the whole image.

10. Composition, Camera, Lighting, and Backgrounds

Anima can understand scene relationships, but it cannot obey details that were never specified. Users often devote fifty tags to a face and only one tag to the entire environment, then interpret the weak background as a model defect.

Camera block

Choose one item from each relevant category:

• Framing: close-up, upper body, cowboy shot, full body, wide shot

• Angle: from above, from below, eye level, from side, three-quarter view

• Lens feeling: wide angle, telephoto, shallow depth of field

• Composition: centered composition, asymmetrical composition, dynamic composition

Avoid stacking mutually exclusive camera angles.

Spatial prose

Tags name objects; prose can place them:

The girl occupies the lower-right foreground. A red torii gate rises behind her at the center of the frame, while the mountain path curves from the lower-left corner into the misty background.

This is more explicit than:

girl, torii, mountain path, mist

Background block

Describe at least:

1. Location

2. One or two defining objects

3. Depth relationship

4. Time or weather

rainy shopping street, glowing shop signs, wet pavement, night.
She stands beneath a transparent umbrella in the foreground. The street recedes diagonally behind her, and warm reflections from the signs stretch across the wet road.

Lighting block

Use physically compatible lighting:

sunset, warm backlighting, rim light, long shadows

or:

overcast daylight, soft diffuse light, low contrast

Do not add every fashionable lighting tag. soft lighting, hard lighting, flat lighting, and dramatic chiaroscuro are different visual requests.

11. Multi-Character Prompting

Multi-character generation remains difficult. Community reports are genuinely mixed: some users find Anima better than SDXL anime checkpoints at interactions and depth placement, while others still experience identity bleed, duplicated subjects, and clothing transfer. It is more accurate to say that Anima's natural-language conditioning gives you better tools for describing the problem—not that it solves the problem automatically.[6][10]

Reliable structure

Start with the scene and count:

safe, 2girls, full body, city park, daytime.
Two women stand side by side on a park walkway and hold opposite ends of the same folded map.

Then define each subject in a separate sentence:

The woman on the left has short black hair and green eyes; she wears a white jacket, black trousers, and red sneakers.
The woman on the right has long blonde hair and blue eyes; she wears a navy cardigan, a yellow skirt, and brown boots.

Then restate the interaction:

The black-haired woman points toward the upper-left corner of the map while the blonde woman looks down at the same point.

Multi-character formula

Count → global scene → left/center/right or foreground/background → identity A → appearance A → identity B → appearance B → interaction → environment

What to avoid

• A comma-only list of two character names

• Describing both wardrobes before assigning them to positions

• Reusing the same color across every subject

• Adding several character LoRAs before the text-only scene works

• Expecting XML, JSON, or bracket labels to act as guaranteed regional conditioning

Users report success with plain sentences such as “the woman on the left…” and “the character on the right…”. Another recurring recommendation is to describe known characters even when the model should recognize them.[6]

If attributes still bleed

1. Simplify to two original characters with one hair color and one clothing color each.

2. Fix the seed.

3. Remove all LoRAs.

4. Use one sentence per subject.

5. Move the two subjects to clearly different positions.

6. Generate a seed grid.

7. Reintroduce the character or style LoRA one at a time.

If the plain model works but the LoRA version fails, the LoRA may have learned unintended pose, costume, or identity correlations.

12. Prompt Weighting and Negative Prompts

Weighting

The official syntax example is:

(chibi:2)

Anima often requires a higher numeric weight than a typical SDXL prompt to create a similar visible push.[1]

Use weights for:

• A concept that repeatedly disappears

• A camera instruction that remains weak

• A style accent

• Resolving competition between two non-contradictory ideas

Do not use weights to repair an incoherent prompt. If full body conflicts with extreme close-up, increasing both weights cannot create a sensible camera.

Controlled weighting ladder

Test:

wide angle shot

then:

(wide angle shot:1.5)

then:

(wide angle shot:2)

Stop when the concept becomes reliable. Stronger is not automatically better.

Negative prompt strategy

Start with the official compact baseline for Base:

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, chromatic aberration

Then add only visible failures:

extra fingers, missing fingers, extra arms

or:

monochrome, grayscale

Do not begin with a hundred-item inherited SDXL negative prompt. Anima is sensitive to negative conditioning, and community users report that oversized negatives can remove detail or introduce unexpected shifts.[5]

Experimental community negatives

Some users report that adding deviantart to the negative prompt can improve the consistency of anime-style outputs. This is plausible because deviantart is an official dataset tag for part of Anima's non-anime artistic data, but the developer does not recommend it as a universal negative. Treat it as an A/B experiment; it may also suppress painterly diversity that you want.[1][5]

Aesthetic version reminder

For Aesthetic, avoid score tags in both positive and negative prompts. A compact negative might be:

worst quality, low quality, artist name, blurry,
jpeg artifacts, chromatic aberration

13. Testing and Troubleshooting

Change one variable

For every prompt experiment, keep constant:

• Model version

• Seed or seed set

• Resolution

• Sampler and scheduler

• Steps

• CFG

• LoRAs

Change only one prompt block.

Use a seed grid

One seed can flatter or punish a prompt by chance. Test at least four seeds before concluding that an artist, quality tag, or negative prompt is reliable.

Diagnostic strip-down

When the image becomes unstable:

1. Remove LoRAs.

2. Remove all weights.

3. Remove all but one artist tag.

4. Remove score tags.

5. Use a short official negative prompt.

6. Reduce the positive prompt to subject, action, camera, and background.

7. Add blocks back one at a time.

This “subtractive debugging” method is repeatedly recommended by users because quality tags, artist tags, and seemingly harmless style words can alter content as well as appearance.[5][9]

Common failures

Symptom

Likely cause

First test

Plain or random Base style

No style, artist, medium, or useful quality guidance

Add one style block or one verified artist

Pony-like flat look

Score tags narrowing the aesthetic

Remove score_7 at fixed seed

Artist has little effect

Missing @, wrong spelling, weak training presence

Use canonical lowercase @artist

Attributes move between characters

Subjects not separately positioned/described; LoRA bleed

Use one sentence per positioned subject and remove LoRAs

Anatomy collapses

Contradictory pose/camera tags or excessive weighting

Remove conflicts and weights

Background is empty

Environment under-specified

Add location, anchors, depth, and light

Image is overcooked

CFG or weights too high

Return to CFG 4–5 and unweighted prompt

Long text is illegible

Model limitation

Use one short word or add text in post

Realism is weak

Outside official model target

Use an appropriate derivative rather than fighting Base

Upscaling: two competing community workflows

Community experience is not settled:

• One group reports that a 1.25× latent upscale followed by roughly 0.25–0.35 denoise cleans up detail.

• Another group prefers generating at a higher base resolution and increasing steps, saying latent upscale is unnecessary or less reliable.

Both can work in a specific graph. Test them from the same initial latent or image rather than presenting either as an Anima law.[5]

Most community threads cited in this guide were written during the Base v1.0 period, before Aesthetic v1.1 and Turbo v1.1 were released. Re-test community recipes on the newer official versions rather than assuming identical behavior.

14. A Small Illustrious-to-Anima Migration Guide

Architecture

SDXL-family checkpoint

Cosmos-Predict2-derived 2B model with Qwen-3 0.6B encoder

Existing assets

SDXL/Illustrious LoRAs and ControlNets

Use Anima-trained assets; compatibility is not automatic

Ordinary tag format

Often copied in Danbooru autocomplete form

Lowercase with spaces; underscores only for score tags

Artist syntax

Depends on checkpoint conventions

Use the mandatory @artist form

Prompt weights

Small SDXL increases are common

Anima may need visibly higher values, such as 1.5–2

Complex relationships

Often handled with tags plus regional tools

First try clear natural-language positions and interactions

Both model families can understand tags and natural language, so “Illustrious is tags-only” would be inaccurate. The practical difference is that Anima was explicitly trained on tags, captions, and mixed tag-caption inputs, and its official workflow uses a different architecture and text encoder.[1][11]

Do not paste an Illustrious prompt, LoRA stack, CFG, sampler, and negative prompt into Anima and judge the architecture from that first output. Translate the prompt into Anima's syntax, reset the settings, and rebuild from the official baseline.

15. Notable Anima Derivatives

This section is intentionally short. The main guide above remains about official Anima.

WAI-ANIMA

WAI-ANIMA is a community checkpoint based on Anima Base 1.0. Its creator recommends a faster 20–30-step, CFG 4–5, Euler A workflow and a conventional quality/negative prefix. It is useful for users who like the WAI aesthetic, but those settings are the derivative author's recipe—not a correction to the official Base documentation.[12]

Nova Anime AM

Nova Anime AM v4 combines Anima Base v1.0 and Aesthetic v1.1. Its release notes emphasize improved prompt handling and stronger aesthetics while retaining the underlying Anima character. Nova also uses its own large quality-tag stack and derivative-specific negative prompt, so prompts copied from Nova galleries should be simplified before testing on official Base.[13]

RDBT Anima

RDBT is a general Anima finetune offered in base and distilled variants. Its creator intentionally avoids baking one dominant default style, recommends explicitly choosing an artist/style, and advises omitting quality tags because its own training data is already quality-filtered. Distilled RDBT supports lower CFG and a different step strategy. This is a useful example of why derivative instructions can legitimately contradict official Base instructions.[14]

Photanima

Photanima deliberately pushes Anima Base toward photography, a domain the official model card says Base is not designed to handle. Current Turbo releases incorporate the official Turbo v1.1 distillation stack and use custom sampler/sigma workflows. It is noteworthy precisely because it changes the target distribution; do not cite its realism success as a capability claim for official Anima.[15]

Anima-2.9B

Anima-2.9B is an unofficial expanded finetune, not a new official CircleStone version. Its creator focuses on later anime knowledge beyond the official September 2025 cutoff and describes the first release as an incomplete preview with roughly 1.7 million training samples and a modern East-Asian anime bias. Its prompt guide strongly emphasizes character-plus-series anchors and detailed appearance. Treat its architecture, knowledge, and settings as a separate branch.[16]

16. Complete Prompt Examples

The following prompts are original templates written for this guide. They are not claimed benchmark results. Generate a seed grid first, then select an image and record the exact version and settings beneath it.

Example 1 — Clean Aesthetic Portrait

Recommended version: Anima-Aesthetic v1.1 Starting settings: 1024×1024, 30–40 steps, CFG 3–4, er_sde

Positive prompt

safe, highres, 1girl, solo, upper body, looking at viewer,
short silver hair, blue eyes, black turtleneck, small silver earrings,
clean lineart, anime coloring, soft cel shading, muted blue palette.

A young woman is framed from the chest upward against a quiet blue-gray studio background. Soft light from the left illuminates her face, while a narrow rim light separates the right side of her hair from the background.

Negative prompt

worst quality, low quality, artist name, blurry,
jpeg artifacts, chromatic aberration, extra fingers

Example 2 — Tag-Only Anime Screenshot

Recommended version: Anima-Base v1.0 Starting settings: 1152×768, 35 steps, CFG 4.5, er_sde

Positive prompt

masterpiece, best quality, score_7, safe, year 2025, highres,
1girl, solo, anime screenshot, official art, school rooftop,
sunset, chain-link fence, sailor uniform, short brown hair,
brown eyes, wind, looking away, profile, upper body,
warm backlighting, lens flare, dramatic sky, anime coloring

Negative prompt

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, chromatic aberration

Suggested A/B test: remove score_7 while keeping the seed fixed. Compare flatness, line density, color, and variation.

Example 3 — Hybrid Fantasy Action Scene

Recommended version: Anima-Base v1.0 Starting settings: 1024×1536, 40 steps, CFG 4.5, dpmpp_2m_sde_gpu

Positive prompt

masterpiece, best quality, safe, newest, highres, absurdres,
1girl, solo, full body, red hair, long braid, green eyes,
white and gold armor, blue cape, longsword, ruined cathedral,
magic circle, flying debris, dynamic composition, from below,
dramatic lighting, volumetric light, clean lineart, detailed background.

A red-haired knight stands in the lower-center foreground and swings her sword upward through a glowing blue magic circle. Broken stone rises around her, the cathedral columns lean into the background, and a shaft of cold light falls through the shattered roof behind her.

Negative prompt

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, chromatic aberration,
extra arms, extra legs, extra fingers, missing fingers

Example 4 — Known Character with Identity Anchors

Recommended version: Anima-Base v1.0 or Aesthetic v1.1 Starting settings: 896×1152, 35 steps, CFG 4 on Base or 3.5 on Aesthetic, er_sde

Positive prompt

safe, highres, 1girl, solo, fern, sousou no frieren,
long purple hair, purple eyes, black coat, white dress,
long sleeves, wooden staff, flower field, overcast daylight,
full body, looking at viewer, calm expression, anime coloring.

Fern from Sousou no Frieren stands slightly left of center in a field of small white flowers. She holds her staff vertically in her right hand; her long purple hair and black coat move gently in the wind, while distant hills fade into the cloudy background.

Negative prompt

worst quality, low quality, artist name, blurry,
jpeg artifacts, chromatic aberration, extra fingers

Example 5 — Two Characters with Separated Attributes

Recommended version: Anima-Base v1.0 Starting settings: 1344×896, 40–45 steps, CFG 4.5, er_sde

Positive prompt

masterpiece, best quality, safe, highres, 2girls,
full body, city park, daytime, detailed background, anime coloring.

Two women stand side by side on a park walkway and hold opposite ends of the same folded map. The woman on the left has short black hair and green eyes; she wears a white jacket, black trousers, and red sneakers. The woman on the right has long blonde hair and blue eyes; she wears a navy cardigan, a yellow skirt, and brown boots. The black-haired woman points toward the upper-left corner of the map while the blonde woman looks down at that same point. Trees and benches recede along the walkway behind them.

Negative prompt

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, extra person,
duplicate, identical twins, extra arms, extra fingers

Diagnostic version: remove all appearance details except hair colors and jacket colors. Confirm positioning first, then add traits back.

Example 6 — Artist-Control Test

Recommended version: Anima-Base v1.0 Starting settings: 1024×1024, 35 steps, CFG 4.5, er_sde, four fixed seeds

Control prompt

masterpiece, best quality, safe, highres, 1girl, solo,
black bob cut, amber eyes, white summer dress, straw hat,
sunflower field, blue sky, upper body, looking at viewer,
gentle smile, clean lineart, anime coloring

Artist test

masterpiece, best quality, safe, highres, 1girl, solo,
@kantoku, black bob cut, amber eyes, white summer dress,
straw hat, sunflower field, blue sky, upper body,
looking at viewer, gentle smile, clean lineart, anime coloring

Weighted artist test

masterpiece, best quality, safe, highres, 1girl, solo,
(@kantoku:1.5), black bob cut, amber eyes, white summer dress,
straw hat, sunflower field, blue sky, upper body,
looking at viewer, gentle smile, clean lineart, anime coloring

Negative prompt

worst quality, low quality, score_1, score_2, score_3,
artist name, blurry, jpeg artifacts, chromatic aberration

This three-way grid is more informative than publishing one attractive cherry-picked seed.

Example 7 — Chibi Weighting Test

Recommended version: Anima-Turbo v1.1 Starting settings: 1024×1024, 10 steps, CFG 1, euler

Positive prompt

safe, highres, 1girl, solo, (chibi:2), tiny body, oversized head,
pink twin tails, star hair clips, yellow raincoat, red rain boots,
jumping over a puddle, city sidewalk, rainy day,
colorful reflections, cheerful expression, clean lineart

Optional negative prompt

worst quality, low quality, blurry, jpeg artifacts,
realistic proportions, extra fingers

At CFG 1, standard classifier-free guidance gives the negative conditioning little or no leverage. Treat this negative as a reusable fallback for workflows that raise guidance; solve most Turbo problems by clarifying the positive prompt first.

Suggested A/B test: compare chibi, (chibi:1.5), and (chibi:2) across the same seeds.

17. Limitations and Final Checklist

Official limitations:

• Realism is not the intended strength of the official model.

• Long or complex rendered text is unreliable; single words and some short phrases may work.

• Short, vague prompts can produce unwanted content.

• Base has a deliberately plain neutral style without aesthetic guidance.[1][2]

Final prompt checklist

☐ I am using an official Anima version or have identified the exact derivative.

☐ My text encoder and VAE match the Anima workflow.

☐ Ordinary tags are lowercase and use spaces, not underscores.

☐ Every artist tag begins with @.

☐ Subject count appears early.

☐ Known characters include the series and visible appearance.

☐ Original characters have a stable identity packet.

☐ Camera instructions do not contradict one another.

☐ The background has a location, anchors, depth, and light.

☐ Multi-character subjects are positioned and described separately.

☐ Weights are used only after the unweighted prompt is coherent.

☐ The negative prompt begins compact and grows only in response to observed failures.

☐ Aesthetic prompts do not automatically inherit Base score tags.

☐ I tested multiple fixed seeds before judging a prompt technique.

Anima rewards specificity, but specificity is not the same as length. The best prompt is not the one with the most tags; it is the one in which every important visual decision has a clear owner. Use tags to define the vocabulary, prose to define the scene, and controlled testing to discover what the model is actually doing.

Thank you for reading, and if this guide helped you, please consider supporting my work by purchasing one of the models on my Tensor.Art profile. Every bit of support motivates me to keep testing new models and creating more in-depth guides for the community. I will comeback soon with the guidebook for Krea