Deepghs MCP

List all public datasets from the DeepGHS organization on HuggingFace.

DeepGHS publishes datasets including Danbooru2024 (8M+ images), Sankaku, Gelbooru, Zerochan, BangumiBase (character frames), site_tags (cross-platform tag database), face/head detection datasets, and more.

Args: params (ListDatasetsInput): - search (Optional[str]): Keyword filter (e.g. 'danbooru', 'character', 'face') - sort (SortBy): Sort by 'downloads', 'likes', 'createdAt', 'lastModified' - limit (int): Results per page, 1–100 (default: 20) - offset (int): Pagination offset (default: 0) - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: Paginated list of datasets with download counts, likes, update dates, tags, and direct HuggingFace links.

List all public models from the DeepGHS organization on HuggingFace.

DeepGHS models include: CCIP (character similarity encoder), WD Tagger Enhanced (anime image tagger with embeddings), aesthetic scorer, anime/real classifier, image type classifier, furry detector, face/head/person detection models, NSFW censor, and style era classifier.

Args: params (ListModelsInput): - search (Optional[str]): Keyword filter (e.g. 'ccip', 'tagger', 'aesthetic', 'face') - sort (SortBy): Sort by 'downloads', 'likes', 'createdAt', 'lastModified' - limit (int): Results per page, 1–100 (default: 20) - offset (int): Pagination offset (default: 0) - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: Paginated list of models with task type, download counts, likes, and links.

List all public Spaces (live demo apps) from the DeepGHS organization on HuggingFace.

DeepGHS spaces include: reverse image search, Danbooru character lookup, anime face/head/person detection demos, CCIP character similarity demo, WD tagger demo, aesthetic scorer demo, and more.

Args: params (ListSpacesInput): - search (Optional[str]): Keyword filter (e.g. 'detection', 'tagger', 'search') - limit (int): Results per page, 1–100 (default: 20) - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: List of Spaces with SDK type, likes, update dates, and direct links.

Get detailed metadata for a specific DeepGHS dataset, model, or space by repo ID.

Returns the full file tree with sizes, all tags, README card metadata, download counts, creation/modification dates, and gating status. Use this before deciding to download — the file tree shows you exactly what tar/parquet files are inside and how large they are.

Args: params (GetRepoInfoInput): - repo_id (str): Full HF repo ID (e.g. 'deepghs/danbooru2024') - repo_type (str): 'dataset', 'model', or 'space' (default: 'dataset') - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: Full repo metadata including file tree with sizes, tags, card data, and a generated cheesechaser download command if applicable.

Schema (JSON mode): { "id": str, "sha": str, "lastModified": str, "tags": list[str], "downloads": int, "likes": int, "cardData": dict, # README metadata "siblings": [ # File tree {"rfilename": str, "size": int, "blobId": str} ], "gated": bool | str }

Search the DeepGHS site_tags dataset — the definitive cross-platform anime tag reference.

The deepghs/site_tags dataset covers 2.5M+ unique tags across 18 platforms: Danbooru, Gelbooru, Pixiv, Sankaku, Wallhaven, Yande.re, Konachan, Zerochan, Rule34, and more. Each tag has category, post count, and aliases per platform.

This is the key tool for the MultiBoru tag normalization problem:

• Danbooru uses snake_case: hatsune_miku

• Zerochan uses Title Case: Hatsune Miku

• Pixiv uses Japanese: 初音ミク This tool maps them all together.

Args: params (SearchTagsInput): - tag (str): Tag in any format/language/platform - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: Cross-platform tag information including canonical names per platform, post counts, tag category (character/copyright/artist/general), and known aliases. Includes direct dataset link for full data access.

Note: This tool returns the HuggingFace dataset info and provides query guidance for the site_tags dataset. For programmatic tag lookup at scale, use the dataset's Parquet or SQLite files directly via the dataset viewer API.

Search for pre-built character image datasets for LoRA training on HuggingFace.

Searches both deepghs (BangumiBase) and CyberHarem namespaces for datasets built around a specific character. These are pre-crawled, pre-cleaned, and ready to use — saving you from having to run waifuc yourself.

CyberHarem datasets are built using the full DeepGHS automated pipeline: crawl → face filter → CCIP identity filter → WD tag → upload to HF.

Args: params (FindCharacterDatasetInput): - character_name (str): Character name to search for (e.g. 'Rem', 'Hatsune Miku') - response_format (ResponseFormat): 'markdown' or 'json'

Returns: str: List of matching character datasets with image counts, sources, download commands, and links. Also suggests waifuc script generation if no pre-built dataset is found.

Generate a ready-to-run waifuc Python script to crawl and clean anime character images for LoRA training.

waifuc is DeepGHS's data pipeline framework. This tool generates a complete, properly-configured script that:

1. Crawls images from the specified sources (Danbooru, Pixiv, Gelbooru, etc.)

2. Converts to RGB and standardizes backgrounds

3. Filters monochrome/sketch/3D images (NoMonochromeAction, ClassFilterAction)

4. Filters duplicate/similar images (FilterSimilarAction)

5. Detects and splits to single-person crops (FaceCountAction, PersonSplitAction)

6. Filters out wrong characters using CCIP AI identity matching (CCIPAction)

7. Tags all images with WD14 tagger (TaggingAction)

8. Crops to target resolution for the specified model format (SD1.5/SDXL/Flux)

9. Exports in the correct format for the target trainer

Crop sizes by model format:

• SD1.5: 512×512 base, bucket range 256–768

• SDXL: 1024×1024 base, bucket range 512–2048

• Flux: 1024×1024 base, bucket range 512–2048

Args: params (GenerateWaifucScriptInput): - character_name (str): Character display name (used in comments/output path) - danbooru_tag (Optional[str]): Danbooru tag e.g. 'rem_(re:zero)' - pixiv_query (Optional[str]): Pixiv search string e.g. 'レム リゼロ' - sources (list[ImageSource]): ['danbooru', 'pixiv', 'gelbooru', 'zerochan', 'sankaku', 'auto'] - model_format (ModelFormat): 'sd1.5', 'sdxl', or 'flux' - content_rating (ContentRating): 'safe', 'safe_r15', or 'all' - output_dir (str): Output directory path - max_images (Optional[int]): Max images to collect - pixiv_token (Optional[str]): Pixiv refresh token (required for Pixiv source)

Returns: str: Complete, ready-to-run Python script with inline comments explaining each pipeline action and its purpose for LoRA training quality.

Generate a cheesechaser Python script to download images from an indexed DeepGHS dataset.

cheesechaser is DeepGHS's tool for selectively downloading images from HuggingFace datasets that are stored as indexed tar archives. Instead of downloading entire multi-GB tar files, you provide a list of post IDs and it extracts only those images.

This is the most efficient way to get specific images from datasets like:

• deepghs/danbooru2024 (~8M images, hundreds of GB total)

• deepghs/gelbooru-webp-4Mpixel (~millions of images)

• deepghs/sankaku_full (~millions of images)

Args: params (GenerateCheesechaserScriptInput): - repo_id (str): HF dataset repo ID (e.g. 'deepghs/danbooru2024') - output_dir (str): Local directory to save downloaded images - post_ids (Optional[list[int]]): Specific post IDs to download - max_workers (int): Parallel download threads (1–16, default: 4)

Returns: str: Complete cheesechaser Python script with inline comments, plus guidance on how to find post IDs from Danbooru/Gelbooru search results.