🌸 AnimeOmni

A richly annotated anime-character dataset for vision-language model training and diffusion-based text-to-image generation.

20,023 labeled samples · updated 2026-08-07 21:36 UTC

---

Overview

AnimeOmni pairs safe-rated anime character images with structured, multi-layered annotations: a detailed natural-language caption, a two-level archetype taxonomy (family → category), a per-character expression label, and the original descriptive tags. Every record also ships pre-formatted as a human/assistant conversation, ready to drop into a vision-language model fine-tuning pipeline with no extra preprocessing.

Unlike a flat caption-only dataset, every sample here is classified along three independent axes — what kind of character, what they're wearing/who they are archetypally, and how they're feeling — so the dataset can be filtered, balanced, or conditioned on any of the three without re-annotating anything.

The dual caption (dense, natural-language) and t2i_tags (short, comma-separated) fields mean the same sample can drive two very different training setups out of the box: vision-language models that need free-form descriptive captions (VQA, image captioning, grounding), and diffusion-based text-to-image generators that need compact, prompt-style tag strings for conditioning.

Archetype Taxonomy

The family / category pair follows a two-level schema. The taxonomy below is illustrative of what's covered — it is a living schema, not a hard-coded restriction, so exact category counts will grow over time.

Family	Example categories
Kemonomimi	kitsune, neko, usagi, ookami, inu, nezumi, hyena, umamusume, tora, hitsuji, kuma
Fantasy & Sci-Fi	dragon_girl, elf, mecha_girl, magical_girl
Horror & Dark	vampire, succubus
Maid & Service	classic_maid, cat_maid, gothic_maid, nurse
Gender Archetypes	bishounen, tomboy, femboy_trap
Outfits & Subcultures	school_uniform, gothic_lolita, kimono_yukata, swimsuit, bunny_suit, casual_hoodie
Everyday & School	(contextual, tag-derived)
Action & Combat	(contextual, tag-derived)
Sci-Fi & Cyberpunk	(contextual, tag-derived)

expression is a free-form but consistent snake_case label describing the character's dominant emotion — common values include embarrassed_blushing, tsundere, smug, cheerful, pouting, serene, neutral, angry_scowling, crying_sadly, surprised_shocked, sleepy, yandere, winking, mischievous, and scared, among others.

For images with multiple characters or expression panels, expression preserves all labels in a single comma-separated string (for example, neutral, cheerful, pouting). This column has the same string type in every record, including multi-expression images, so JSONL loading and the dataset viewer use a consistent schema.

Dataset Structure

The dataset is released in incremental batches. Each top-level folder is a contiguous ID range, e.g. 000001-000500:

000001-000500/
  metadata.jsonl
  data/
    image_12.jpg
    image_57.jpg
    ...

Schema

Each line of metadata.jsonl is one JSON record:

Field	Type	Description
pid	int	Source post identifier
file_name	string	Relative path to the image (data/...)
caption	string	Detailed description of the character, outfit, pose, and scene
family	string	Top-level archetype family
category	string	Specific archetype in snake_case
expression	string	Dominant character emotion in snake_case
tags	list[string]	Descriptive tags for the character/scene
t2i_tags	string	Same tags, comma-separated, ready for T2I conditioning
width, height	int	Image dimensions after normalization
aspect_ratio	float	width / height, rounded
rating	string	Content rating (safe)
source	string	Origin identifier
vlm_conversation	list[dict]	human/gpt turns, ready for chat-style VLM fine-tuning

Example record

{
  "pid": 184213,
  "file_name": "data/image_184213.jpg",
  "caption": "A female character with long silver hair tied in a high ponytail and amber eyes, wearing a fox-eared hood over a dark kimono-style jacket with a red obi. She stands in a three-quarter pose against a soft autumn background, with a gentle, closed-mouth smile.",
  "family": "Kemonomimi",
  "category": "kitsune",
  "expression": "serene",
  "tags": ["fox_ears", "fox_tail", "silver_hair", "kimono", "autumn", "smile"],
  "t2i_tags": "fox_ears, fox_tail, silver_hair, kimono, autumn, smile",
  "width": 832,
  "height": 1024,
  "aspect_ratio": 0.81,
  "rating": "safe",
  "source": "imageboard",
  "vlm_conversation": [
    {"from": "human", "value": "<image>\nDescribe the anime character appearance, outfit, and expression in detail."},
    {"from": "gpt", "value": "A female character with long silver hair tied in a high ponytail..."}
  ]
}

(Illustrative example — not a literal dataset row.)

Annotation Process

Captions, archetype labels, and expressions are produced by an AI vision-language model, grounded on each image together with its descriptive tags to reduce misidentification of species, gender, and visible traits. A deterministic tag→category override is applied for a curated set of unambiguous archetypes (e.g. specific kemonomimi ear/tail tags), so those labels stay consistent rather than purely model-inferred.

Quality Control

Every candidate image and annotation passes through automated checks before being added to the dataset:

• Resolution & framing — minimum dimensions and aspect-ratio bounds are enforced.
• Blank/placeholder detection — low-contrast, near-empty, or degenerate images are rejected.
• Duplicate detection — exact duplicates are caught via cryptographic hashing of image content; near-duplicates (re-uploads, re-compressions) are caught via perceptual hashing, with a persistent index maintained across the whole dataset.
• Tag sufficiency — samples with too few descriptive tags are skipped, since they provide insufficient grounding for annotation.
• Caption validation — captions that are too short, empty, or resemble a non-answer are discarded and the sample is retried or dropped.
• Oversized-source rejection — images whose declared pixel dimensions exceed a format-aware safety threshold are rejected before decoding, to protect against decompression-bomb style inputs.
• Post-publication integrity checks — every batch is verified immediately after upload, and re-verified on subsequent runs; any batch where the image count doesn't match its metadata is automatically removed.

Intended Uses

Vision-language / multimodal models

• Fine-tuning or evaluating VLMs on anime-style character captioning, using vlm_conversation directly.
• Image-to-text and visual question-answering research grounded in caption.

Diffusion / text-to-image generation

• Fine-tuning or LoRA-training diffusion-based image generators, using t2i_tags (short, prompt-style) or caption (dense, descriptive) as the text conditioning.
• Tag-to-image and archetype-conditioned generation, using family/category/expression to steer style, subject, or mood.
• Building curated, style-balanced training subsets for character-generation models.

Classification & retrieval

• Archetype and expression classification research.
• Building balanced training subsets by filtering on family, category, or expression.
• Text-to-image and image-to-image retrieval benchmarking using the tag/caption pairs.

Limitations & Biases

• Captions and taxonomy labels are AI-generated and, despite tag grounding and validation, may contain occasional inaccuracies or inconsistent granularity between samples.
• The archetype taxonomy reflects a fixed, curated schema and may not capture every visual nuance of a given character.
• Source images are drawn from a single imageboard-style platform and may not represent the full stylistic diversity of anime art.

License & Attribution

This dataset's compiled structure, and the annotations layer (captions, taxonomy, expression labels, and metadata) created for it, are released under CC BY 4.0. Attribution should reference this dataset and repository.

Underlying images originate from third-party artists; this license covers the annotation and compilation work, not any pre-existing rights in the original artwork itself. Please verify suitability for your use case before redistribution.

If you're using AnimeOmni in your research, LoRA, or project, feel free to open a Discussion or ping me, I’d love to see what you’re building!