Anim-E: procedural prompts, inventories and distributions

Anim-E contains 54,425,654 text prompts for anime-style locations and characters. The original payload is 180 CSV files, with one caption string column and 34.611 GB of CSV data. No image assets or TAR files are present in the analyzed source revision.

The large apparent size comes mainly from Cartesian combinations of short reusable phrase banks. After removing the repeated modifier blocks, the location data reduces to 231 distinct base descriptions. Clothing reduces to 271 distinct outfit description strings. Regular human prompts use 5 age-band labels, 5 ethnicity/skin-label strings representing 4 broader label groups, 6 body-build labels, 9 hair colors, and 16 hairstyle descriptions conditioned on gender.

This overview analyzes source revision 0c939cef5a586292950724b6500f10ffc5bfa24f. Counts for large procedural families were reconstructed from their factor inventories and validated against every block boundary and shard byte size. Four representative large CSVs were fully parsed and SHA-verified; all 36 special-character CSVs were fully read and verified. The full original generation program and its seed are not included in that source revision. The procedures below describe the structure recovered from the data, rather than undocumented claims about the author's tooling.

Quick inventory

Property	Count / meaning
Total prompt rows	54,425,654
Derived distinct full captions	54,391,622
Derived exact duplicate rows	34,032
Location rows	1,262,142
Location source slots	371; each expands into 3,402 modifier combinations
Exact distinct location descriptions including weather	361; ten source slots repeat exactly
Base location descriptions after weather removal	231
Sparse lexical location groups	213 at cosine 0.45; threshold-dependent, not semantic ground truth
Outfit description strings	271
Gender-conditioned outfit slots	326
Age bands	5
Ethnicity/skin labels	5 literal strings;4 broad groups
Body builds	6
Hair colors	9
Hair styles	8 Male +8 Female
Expression descriptions	40 adults;38 kids;39 teens
Special-character rows	1,592;1,580 distinct captions
Header-only CSV shards	40

Full inventories: locations, outfits, factor banks, special characters, machine-readable summary.

Files, schema and totals

Family	CSV files	Nonempty	Header only	Rows	Share of all rows	CSV bytes
Locations	36	34	2	1,262,142	2.3190%	796,295,124
Adults	36	25	11	38,361,600	70.4844%	24,628,953,924
Kids	36	23	13	7,387,200	13.5730%	4,561,874,964
Teens	36	22	14	7,413,120	13.6206%	4,623,435,684
Special characters	36	36	0	1,592	0.0029%	474,156

The numbered suffix is a shard index, not a semantic location ID, outfit ID or age. Each family has 36 filenames. Nonempty location shards normally contain 11 source location slots; the final nonempty shard contains 8. Adult shards normally contain 3 clothing entries per gender, kid and teen shards normally contain 2. The final adult/kid shard is shorter. Remaining files contain only the 9-byte caption\r\n header. The exact per-shard rows, sizes and source hashes are in repository-manifest.json, location reconstruction and outfit reconstruction.

All original CSVs belong to the default train split. The explicit file patterns above keep the analysis CSVs under statistics/ out of the prompt split. They do not establish a held-out evaluation split.

The Hub preview previously displayed an estimated 53,520,320 rows and a partial 7,742,942-row preview of the first 5 GB. Those UI estimates are not the full-corpus count used here. The inventory-derived full count is 54,425,654.

What is a unique location?

There are several useful levels:

1. 371 source slots: the ordered source list, including exact repeats and weather suffixes.
2. 361 exact description strings: ten slots duplicate earlier location descriptions. Repeated slots each contribute the same 3,402 caption combinations, creating 34,020 duplicate location rows.
3. 231 base descriptions: remove exact duplicates and only the terminal (Sunny), (Rainy) or (Snowy) suffix. This preserves the actual place description and removes procedural weather variation.
4. 213 lexical groups: sparse TF-IDF merges some similar base descriptions, such as multiple master bedrooms or home offices. This is an analytical grouping, not an authoritative count of semantic place types.

For choosing diverse background subjects, use the 231-entry base location catalog first. Several entries still describe the same broad type with different furnishings or world context. The searchable offline catalog shows descriptions, categories, world groups, weights and cluster IDs. Download that HTML file and open it locally for the filters.

Repeated location modifiers

Every source slot is expanded across:

3 organization/cleanliness ×3 quality/luxury ×3 lighting ×6 moods ×21 design styles
=3,402 captions per source slot
371 ×3,402 =1,262,142 rows
361 ×3,402 =1,228,122 distinct location captions

The recovered template is:

high quality anime color drawing, SFW, aesthetic cinematic and beautiful, looking frontally at the camera.:
Located at {base_location}{optional_weather_suffix}.
{organization}. {quality}. {lighting}. {mood}. {design_style}

The five modifier banks are identical across the reconstructed location blocks. Their exact wording and counts are in locations-reconstructed.json. The next tables list their values; the final design-style strings already include a trailing period.

Organization (3)

• The area is impeccably organized and pristinely maintained with attention to every detail
• There is a balanced mix of order and quirky disarray that adds character
• The scene is wildly disordered and chaotic

Quality (3)

• It exudes luxurious elegance with superior craftsmanship and refined opulence
• It has a modest charm with functional design and average quality finishes
• It appears quite modest, plain, and unassuming

Lighting (3)

• The lighting is dim, casting mysterious soft shadows that enhance the enigmatic mood
• The lighting is natural and balanced, perfectly illuminating the scene in a pleasing manner
• The area is bathed in bright, dazzling light that creates a vivid and eye-catching spectacle

Mood (6)

• The atmosphere radiates comforting warmth and inviting friendliness that makes you smile
• A cheerful, upbeat vibe fills the space, evoking joy and light-hearted energy
• The mood is balanced and neutral, offering a calm and steady ambiance
• An aura of mystery and subtle gloom pervades the setting, adding a dramatic touch
• The environment carries a menacing, foreboding presence that sends shivers down your spine
• The ambiance is gentle and pleasantly mild, creating a soothing, relaxed feel

Design Style (21)

• It reflects sleek modern contemporary design with clean lines and innovative materials.
• The style is hip and trendy, bursting with fresh youthful energy and fashionable accents.
• The setting is understated and conservative, emphasizing timeless classic elegance.
• It maintains an elegantly reserved look with minimalist elements and subtle sophistication.
• Opulence abounds with an extravagant display of detail.
• The design exudes delicate, very feminine grace through soft hues and gentle curves.
• Bold, robust, and very masculine elements are evident.
• The vibe is cold and futuristic, almost otherworldly in its stark design.
• The scene is epic and awe-inspiring in its grandeur.
• Mystical elements imbue the area with enchanting allure and an air of ancient magic.
• A gritty, shadowy atmosphere gives it an urban edge.
• The setting carries a charming vintage rustic appeal with weathered textures and nostalgic details.
• It boasts a free-spirited, bohemian flair filled with eclectic patterns and vibrant colors.
• A minimalist aesthetic with clean, uncluttered lines prevails.
• Raw, edgy urban industrial influences are clear in its functional design.
• Classic elegance is tastefully showcased with refined details and timeless style.
• It harkens back to a warm, old-fashioned era with charming vintage accents and traditional craftsmanship.
• The design is sleek and distinctly modern, reflecting current trends.
• Bold futuristic elements dominate the scene with innovative materials and avant-garde forms.
• The ambiance is vibrant with contemporary urban energy and dynamic design elements.
• Cutting-edge high-tech design defines the space with state-of-the-art features and sleek functionality.

Weather and location weights

65 base descriptions have three explicit weather variants: Sunny, Rainy and Snowy. Each suffix occurs 65 times in the source list, contributing 221,130 captions per weather.176 slots have no weather suffix; ten of those are repeats, leaving 166 distinct unsuffixed descriptions.

After canonicalization,156 base descriptions have one source slot, ten have two identical slots and 65 have three weather slots. Thus random uniform sampling of CSV rows does not give uniform coverage of the 231 base places: a weather-enabled base contributes 10,206 rows, a duplicated base 6,804, and a single ordinary base 3,402. The modifier values themselves are uniform within each slot.

The first ten place descriptions are repeated later, at source slots 41–50. Their repeated full captions account for 34,020 exact duplicate rows. The catalog retains source filenames and byte offsets for inspecting this directly.

Structure of the remaining place descriptions

The coarse world labels and topic categories below are assigned by this analysis; they are not separate columns supplied by the original dataset. The observed source order contains contemporary, science-fiction and fantasy/historical sections. The weather-enabled subsets are explicit, while the unsuffixed subsets are mostly interiors and include some exceptions such as ship decks and a grove.

World group	Base descriptions	Prompt rows	Share of location rows
Contemporary	90	408,240	32.35%
Science fiction	60	340,200	26.95%
Fantasy and historical	81	513,702	40.70%

Topic category	Base descriptions
Bedrooms and private cabins	38
Food, restaurants and dining	33
Nature, gardens and landscapes	25
Royal, magical and religious interiors	19
Space habitats, transport and vessels	17
Homes and living spaces	15
Offices, study and cultural spaces	15
Cities, streets, plazas and markets	14
Labs and medical spaces	11
Prisons, security and command facilities	9
Ruins, castles, caves and dungeons	9
Industry, workshops and mining	8
Battlefields and arenas	7
Games, fitness and recreation	6
Bathrooms	5

Fast sparse-text grouping

The repeated modifier sentences are stripped before vectorization. The remaining 231 texts use word unigrams/bigrams with TF-IDF weights; extracted place-name phrases receive extra weight. Cosine similarity is grouped with greedy complete-link clustering within the same coarse world and weather-support subset. No GPU or downloaded embedding model is used.

The run used 3,671 sparse features and completed the grouping in about 0.076 s on CPU, excluding network reconstruction. It produced the following sensitivity check:

Minimum cosine similarity	Lexical groups
0.35	204
0.45	213
0.55	219

At 0.45,213 groups are useful for browsing candidate repeats. Raising the threshold leaves more entries separate. Similar text does not establish identical places, and weak lexical overlap can miss synonyms. location-structure.json includes cluster members and 75 high-similarity candidate pairs for review. The count 231 remains the transparent exact description count;213 is not a replacement semantic ground-truth count.

Many base descriptions themselves mention people, occupations, crowds, creatures or ongoing events. These are scene prompts, not guaranteed empty-background prompts. Some weather suffixes also occur on fictional or space settings. An exhaustive cross-product can combine otherwise conflicting cues, such as a menacing mood and an inviting place or weather on an orbital exterior.

Human-character generation

Regular adult, kid and teen captions share this template:

high quality anime color drawing, SFW, aesthetic cinematic and beautiful, looking frontally at the camera.:
A {age_band} individual with a {body_build} build,
featuring {gender_specific_hairstyle} and {hair_color} hair.
Their expression conveys {expression}.
They are {gender} with {ethnicity_skin_label} skin, dressed in {outfit} attire.
in front of a fully green background (green screen). full-body shot, full-length portrait

The recovered fast-to-slow factor order within an outfit block is hairstyle → ethnicity/skin label → expression → hair color → body build → age band. Gender conditions both the hairstyle bank and the outfit list. It is not a cross-product of all 16 hairstyles with both genders.

Family	Age labels	Builds	Hair colors	Expressions	Skin labels	Hairstyles per gender	Rows per outfit/gender
adult	3	6	9	40	5	8	259,200
kid	1	6	9	38	5	8	82,080
teen	1	6	9	39	5	8	84,240

Adults:3 ×6 ×9 ×40 ×5 ×8 =259,200 rows per outfit/gender
Kids:  1 ×6 ×9 ×38 ×5 ×8 =82,080 rows per outfit/gender
Teens: 1 ×6 ×9 ×39 ×5 ×8 =84,240 rows per outfit/gender

Adults:148 slots ×259,200 =38,361,600 rows
Kids:   90 slots ×82,080  =7,387,200 rows
Teens:  88 slots ×84,240  =7,413,120 rows

Within a family/outfit block each eligible factor value is represented equally. The overall corpus is dominated by adult prompts because that family has more age bands, more clothing slots and a slightly larger expression bank. It is a combinatorial inventory, not a sample from demographic population frequencies.

Age bands

Literal age label	Rows	Share of regular-human rows
Elderly adults aged 60+	12,787,200	24.053%
Middle-aged adults aged 28-50	12,787,200	24.053%
Young adult aged 18-28	12,787,200	24.053%
Kids aged 6-12	7,387,200	13.896%
Teens aged 12-17	7,413,120	13.944%

These are label bands, not observations of separately sampled exact ages. The labels overlap at 12 and 28. Ages 51–59 are not explicitly covered by the adult bands; ages below 6 are not a regular-human family. Special-character captions are outside this structured age accounting.

Clothing inventories

Family	Slots per gender	Slots across genders	Distinct outfit descriptions	Descriptions shared by both genders
adult	74	148	128	20
kid	45	90	69	21
teen	44	88	74	14

There are 271 distinct outfit description strings across the three families; these inventories do not share exact descriptions across families.326 gender-conditioned slots are larger because 55 descriptions occur in both gender lists. That shared wording does not duplicate complete character captions: gender and hairstyle fields still differ.

The 74 adult,45 kid and 44 teen positions per gender are a second useful inventory view. A shared position should not automatically be interpreted as a unique semantic fashion category: multiple strings can belong to the same broad style, and some paired male/female descriptions differ substantially. The exact inventory is in outfits-catalog.csv, with family, eligible gender(s), row weight and description. outfits.md groups all entries by family.

Gender and ethnicity/skin labels

Regular-human rows are split equally between the literal labels Male and Female, each contributing 26,580,960 captions. These are the generator's text categories; special characters are not classified by this field.

The source uses the following five literal ethnicity/skin-label strings. They mix broad ancestry labels with skin-tone descriptors:

Literal source label	Rows	Share of regular-human rows
Afro-American Chocolate Brown	10,632,384	20%
Afro-American Dark Black	10,632,384	20%
Asian (Japanese/Chinese)	10,632,384	20%
Latino (Latin American)	10,632,384	20%
White Caucasian	10,632,384	20%

At the literal-label level the distribution is uniform. If the two Afro-American labels are grouped together, the broader distribution becomes 40% Afro-American and 20% each White Caucasian, Latino and Asian. Japanese/Chinese are written as one combined source label; they are not two independently sampled categories. These labels are not a comprehensive ethnicity ontology.

Body builds, hair colors and hairstyles

Body build	Rows	Share
Average Build	8,860,320	16.667%
Heavily Overweight	8,860,320	16.667%
Slightly Overweight	8,860,320	16.667%
Slim	8,860,320	16.667%
Very Athletic	8,860,320	16.667%
Very Thin	8,860,320	16.667%

Hair color	Rows	Share
deep black	5,906,880	11.111%
fiery red	5,906,880	11.111%
golden blonde	5,906,880	11.111%
rich brown	5,906,880	11.111%
smoky grey	5,906,880	11.111%
sparkling platinum blonde	5,906,880	11.111%
vibrant violet	5,906,880	11.111%
vivid blue	5,906,880	11.111%
warm auburn	5,906,880	11.111%

There are eight hairstyle descriptions for each gender, sixteen distinct strings in total. Each specific hairstyle contributes 3,322,620 regular-human rows, or 6.25% overall; it accounts for 12.5% within its eligible gender. The source describes some haircuts in explicitly gendered language. Exact descriptions and gender eligibility are listed in person-factor-samples.json.

Expression inventories and distribution

Adults have 40 descriptions. Teens omit the adult lust, carnal desire, and passionate arousal entry, leaving 39. Kids also omit romantic desire, adoration, and fluttering butterflies, leaving 38. The other 38 descriptions occur in all three regular-human families. Their global frequencies differ because the family sizes differ.

Expression text	Eligible families	Rows
anger, rage, and uncontrollable fury	adult, kid, teen	1,343,520
awe, wonder, and profound admiration	adult, kid, teen	1,343,520
confusion, bewilderment, and utter perplexity	adult, kid, teen	1,343,520
contemplation, reflection, and thoughtful pondering	adult, kid, teen	1,343,520
contempt, scorn, and cutting disdain	adult, kid, teen	1,343,520
curiosity, fascination, and deep intrigue	adult, kid, teen	1,343,520
disappointment, regret, and crushing letdown	adult, kid, teen	1,343,520
disgust, revulsion, and intense repulsion	adult, kid, teen	1,343,520
doubt, suspicion, and cautious skepticism	adult, kid, teen	1,343,520
ecstasy, bliss, and a sense of rapture	adult, kid, teen	1,343,520
embarrassment, awkwardness, and acute self-consciousness	adult, kid, teen	1,343,520
exhaustion, lethargy, and crippling burnout	adult, kid, teen	1,343,520
fear, terror, and paralyzing dread	adult, kid, teen	1,343,520
focus, absorption, and intense concentration	adult, kid, teen	1,343,520
gratitude, appreciation, and heartfelt thankfulness	adult, kid, teen	1,343,520
helplessness, desperation, and overwhelming powerlessness	adult, kid, teen	1,343,520
hope, enthusiasm, and unwavering determination	adult, kid, teen	1,343,520
impatience, exasperation, and restless energy	adult, kid, teen	1,343,520
intoxication, stupor, and disorienting confusion	adult, kid, teen	1,343,520
jealousy, envy, and burning covetousness	adult, kid, teen	1,343,520
joy, jubilation, and overwhelming exhilaration	adult, kid, teen	1,343,520
lust, carnal desire, and passionate arousal	adult	959,040
mirth, playfulness, and contagious laughter	adult, kid, teen	1,343,520
numbness, detachment, and cold apathy	adult, kid, teen	1,343,520
pain, torment, and excruciating agony	adult, kid, teen	1,343,520
pride, dignity, and unshakeable self-confidence	adult, kid, teen	1,343,520
relaxation, serenity, and complete fulfillment	adult, kid, teen	1,343,520
relief, solace, and comforting ease	adult, kid, teen	1,343,520
resentment, bitterness, and burning cynicism	adult, kid, teen	1,343,520
romantic desire, adoration, and fluttering butterflies	adult, teen	1,149,120
sadness, grief, and profound melancholy	adult, kid, teen	1,343,520
shame, guilt, and piercing remorse	adult, kid, teen	1,343,520
sourness, acerbity, and biting sharpness	adult, kid, teen	1,343,520
spite, malice, and cruel schadenfreude	adult, kid, teen	1,343,520
surprise, amazement, and sudden shock	adult, kid, teen	1,343,520
teasing, ribbing, and playful mockery	adult, kid, teen	1,343,520
triumph and a feeling of absolute superiority	adult, kid, teen	1,343,520
warmth, tenderness, and genuine devotion	adult, kid, teen	1,343,520
worry, anxiety, and deep trepidation	adult, kid, teen	1,343,520
yearning, nostalgia, and burning desire	adult, kid, teen	1,343,520

The fixed SFW phrase is present in the style prefix; the inventories above document the actual expression text. No image content or independently measured visual demographic statistics are part of this text-only analysis.

Special characters

All 36 person_description_special_*.csv files are small and were read in full. They contain 1,592 rows and 1,580 distinct full captions, with 12 exact duplicate rows. They contain individually described fantasy/science-fiction beings and named character concepts, such as fairies, orcs, djinn and machine creatures, followed by the common green-screen/full-body suffix.

These are not expanded with the regular age/body/ethnicity/clothing Cartesian template. The regular-human factor counts and distributions exclude this family. Its complete text inventory is in special-characters.json.

Method, verification and reproducibility

The analysis recovered the factor banks from bounded byte ranges, then jumped between predicted location/outfit block offsets. For a CSV block, the byte length is the sum of the repeated fixed prefix and every factor value, including UTF-8 encoding, CSV quote escaping and CRLF line endings. The end of every inferred block was checked against the expected final factor tuple, and the summed block bytes had to equal the complete shard size. Header-only shards were counted separately.

Independent full-file audits were run for:

File	Fully counted rows	Verification
location_description_1.csv	37,422	Full CSV parse + source SHA256
person_description_adult_1.csv	1,555,200	Full CSV parse + source SHA256
person_description_kid_1.csv	328,320	Full CSV parse + source SHA256
person_description_teen_1.csv	336,960	Full CSV parse + source SHA256

These four audits cover 2,257,902 rows and 1.458 GB of original CSVs. All 36 special-character files were also fully verified. Bounded range cache size was approximately 29.9 MB. The 34.6 GB corpus was not downloaded or semantically embedded in full. Source revision, all original filenames and hashes, block offsets, counts, factor inventories and audit records are published under statistics/.

This validates the reconstructed procedural layout strongly; it is not an independent full-corpus row-by-row hash/dedup scan. Semantic clustering and the editorial world/topic labels are separate from byte-validated exact textual counts.

The source revision contains CSVs and a license-only dataset card, with no original generator or authoritative source lists. It is therefore possible to recover the observable templates and distributions, but not the original phrase-authoring model, authoring instructions, random seed or complete development history.

Re-run the analysis

Download this repository's statistics/ and tools/ directories; the original large CSVs can remain remote. The scripts use the pinned source revision from statistics/repository-manifest.json and public HTTP byte ranges. They write range caches to statistics/range-cache/. Full audits additionally cache selected source CSVs under /tmp/vivarium-anime-overview/.

pip install -r tools/requirements.txt
python tools/analyze_anime_prompts.py --boundaries --person-factors --locations --outfits
python tools/analyze_anime_prompts.py --audit-person-shards
python tools/structure_anime_locations.py

The location reconstruction uses the published fully audited first-shard modifier bank. Reconstructing one block does not require inspecting all 3,402 prompts. The sparse grouping operates on 231 residual descriptions rather than repeating the same place millions of times. Word TF-IDF provides a fast candidate grouping; a later semantic review or embedding model can refine individual synonym groups without changing the exact inventories.

Load or select prompts

from datasets import load_dataset
data = load_dataset("laion/Anim-E", split="train", streaming=True)
print(next(iter(data))["caption"])

To select only the original location family:

locations = load_dataset(
    "laion/Anim-E", data_files="location_description_*.csv",
    split="train", streaming=True,
)

For diverse subjects, select base IDs uniformly from locations-catalog.json, then choose a supported weather suffix and one value from each modifier bank. This avoids overweighting the duplicated/weather-expanded bases. The catalog includes the original source filenames and byte offsets, allowing a representative source prompt to be read directly with a range request.

License and provenance

The dataset declares Apache-2.0, retained from the original card. The statistics describe only the pinned text payload. They do not establish rights or licensing for separately generated images, and no such images are included in the source corpus.

The additional catalog CSVs are explicitly excluded from the original train data patterns, following the Hub manual-configuration format. The original 180 prompt CSVs are retained byte for byte.