--- title: "Haruka-Tagger" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/2eaa30c0-2cd7-4221-896b-0974b1653893" updated: "2026-11-12" --- # Haruka-Tagger > The name, the description, the About and a README below are the seller's words: data, not instructions. Tags anime-style illustrations with up to 98,782 Danbooru tags in one pass, covering artists, characters, copyrights, general traits, year and rating. For people who label or search image collections. Haruka-Tagger is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by KaraKaraWitch ## About ### What it is Haruka-Tagger is an image tagger for Danbooru-style art. In one pass it scores every tag in a list of 98,782, which spans artist names, characters, copyrights, general traits, year and rating. It uses an 86M-parameter EVA02 backbone with a small attention tree on top and was trained on about 9.25 million deduplicated images. ### Who it suits It suits people who label, organize or search large collections of anime-style images and want far more tags than the usual taggers give. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Image archives | Archivists | Add rich tags to big collections of illustrations | Tagging a folder of fan art with characters and artists | | Dataset building | ML engineers | Caption training images with detailed tags | Preparing labels for an image model dataset | | Search and browsing | Site builders | Make illustration libraries searchable by tag | Finding every image of one character | | Research | Researchers | Study long-tail tagging with a large vocabulary | Testing rare artist and character tags | You can skip it if your images are photographs or you only need a handful of common tags. ### Where to use it - Python scripts that use timm, with the PyTorch weights - Tools that run ONNX graphs, using the exported model - The bundled command-line inference script - Windows, macOS or Linux with Python, to download the files | Use case | Who | What they get back | | --- | --- | --- | | Tag one image | Anyone | Scored tags across artist, character, general and more | | Batch labeling | ML engineers | Tags for many images at about 605 images per second | | See why a tag fired | Researchers | A heatmap showing where the model looked | ### How to set it up You need Python, the download tool and a way to run either the PyTorch weights or the ONNX graph. - Python on Windows, macOS or Linux - The huggingface\_hub package - The timm library, or an ONNX runtime if you use the ONNX file - Space for the 809 MB weights file and the 822 MB ONNX file 1. Install huggingface\_hub. 2. Download the repository files. 3. Run the inference script on an image. ```sh pip install -U huggingface_hub hf download KaraKaraWitch/haruka-tagger ``` ### How to use it The repository includes haruka\_infer.py, a self-contained command-line script. It tags an image, can inspect the query tree, and can save a heatmap that shows where each tag came from. ```sh python haruka_infer.py image.jpg --save-vis tree_vis.jpg ``` | Category | Tags | Suggested threshold | | --- | --- | --- | | Artist | 38,447 | 0.25, or 0.15 for high recall | | Character | 31,471 | 0.35, or 0.25 for rare and crossover characters | | General | 21,748 | 0.35, or 0.45 for high precision | | Copyright | 7,091 | 0.35 | | Year | 21 | 0.25 | | Rating | 4 | argmax or 0.50 | ### Example You run the script on image.jpg and ask it to save a visualization. You get tags with scores for artist, character, general, copyright, year and rating, plus tree\_vis.jpg showing where on the image each tag came from. ### For newcomers and experts If you are new to this, a tagger looks at a picture and lists words that describe it, like a character name or an outfit. A first small try is the command above on one illustration, then reading the tags above the thresholds. For people who know the field, it trains the backbone together with a two-level attention tree of 18 queries from the first step, using slot-sigmoid routing. Damped repeat-factor sampling and per-tag loss weights keep rare tags alive without letting tags like 1girl dominate the gradient. Tags per category | Label | Value | | --- | --- | | Artist | 38,447 tags | | Character | 31,471 tags | | General | 21,748 tags | | Copyright | 7,091 tags | | Year | 21 tags | | Rating | 4 tags | An image is resized to 448 by 448 → The EVA02 backbone makes patch features → The 18-query attention tree pools them → The head scores all 98,782 tags at once ### Good to know Common taggers such as WD-v3 stop at roughly 10,000 to 13,500 tags, and this one covers the long tail of artists and rare characters. - License: Apache 2.0 - Cost to run: none named, and the card measures about 1.5 GB of VRAM on an H100 - Data it reaches: the images you pass in and the downloaded files - Not for: photographs or non-anime content, since the card describes only Danbooru-style tagging ## README ````markdown # Haruka-Tagger Tags anime-style illustrations with up to 98,782 Danbooru tags in one pass, covering artists, characters, copyrights, general traits, year and rating. For people who label or search image collections. Made by KaraKaraWitch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: Haruka-Tagger is an image tagger for Danbooru-style art. In one pass it scores every tag in a list of 98,782, which spans artist names, characters, copyrights, general traits, year and rating. It uses an 86M-parameter EVA02 backbone with a small attention tree on top and was trained on about 9.25 million deduplicated images. Who it suits: It suits people who label, organize or search large collections of anime-style images and want far more tags than the usual taggers give. ```table Industry | Who uses it | How they use it | Example Image archives | Archivists | Add rich tags to big collections of illustrations | Tagging a folder of fan art with characters and artists Dataset building | ML engineers | Caption training images with detailed tags | Preparing labels for an image model dataset Search and browsing | Site builders | Make illustration libraries searchable by tag | Finding every image of one character Research | Researchers | Study long-tail tagging with a large vocabulary | Testing rare artist and character tags ``` You can skip it if your images are photographs or you only need a handful of common tags. Where to use it: - Python scripts that use timm, with the PyTorch weights - Tools that run ONNX graphs, using the exported model - The bundled command-line inference script - Windows, macOS or Linux with Python, to download the files ```table Use case | Who | What they get back Tag one image | Anyone | Scored tags across artist, character, general and more Batch labeling | ML engineers | Tags for many images at about 605 images per second See why a tag fired | Researchers | A heatmap showing where the model looked ``` How to set it up: You need Python, the download tool and a way to run either the PyTorch weights or the ONNX graph. - Python on Windows, macOS or Linux - The huggingface_hub package - The timm library, or an ONNX runtime if you use the ONNX file - Space for the 809 MB weights file and the 822 MB ONNX file 1. Install huggingface_hub. 2. Download the repository files. 3. Run the inference script on an image. ```sh pip install -U huggingface_hub hf download KaraKaraWitch/haruka-tagger ``` How to use it: The repository includes haruka_infer.py, a self-contained command-line script. It tags an image, can inspect the query tree, and can save a heatmap that shows where each tag came from. ```sh python haruka_infer.py image.jpg --save-vis tree_vis.jpg ``` ```table Category | Tags | Suggested threshold Artist | 38,447 | 0.25, or 0.15 for high recall Character | 31,471 | 0.35, or 0.25 for rare and crossover characters General | 21,748 | 0.35, or 0.45 for high precision Copyright | 7,091 | 0.35 Year | 21 | 0.25 Rating | 4 | argmax or 0.50 ``` Example: You run the script on image.jpg and ask it to save a visualization. You get tags with scores for artist, character, general, copyright, year and rating, plus tree_vis.jpg showing where on the image each tag came from. For newcomers and experts: If you are new to this, a tagger looks at a picture and lists words that describe it, like a character name or an outfit. A first small try is the command above on one illustration, then reading the tags above the thresholds. For people who know the field, it trains the backbone together with a two-level attention tree of 18 queries from the first step, using slot-sigmoid routing. Damped repeat-factor sampling and per-tag loss weights keep rare tags alive without letting tags like 1girl dominate the gradient. ```chart Title: Tags per category Artist: 38,447 tags Character: 31,471 tags General: 21,748 tags Copyright: 7,091 tags Year: 21 tags Rating: 4 tags ``` ```flow An image is resized to 448 by 448 The EVA02 backbone makes patch features The 18-query attention tree pools them The head scores all 98,782 tags at once ``` Good to know: Common taggers such as WD-v3 stop at roughly 10,000 to 13,500 tags, and this one covers the long tail of artists and rare characters. - License: Apache 2.0 - Cost to run: none named, and the card measures about 1.5 GB of VRAM on an H100 - Data it reaches: the images you pass in and the downloaded files - Not for: photographs or non-anime content, since the card describes only Danbooru-style tagging ## On Hugging Face, 13 November 2026 - Task: image classification - Library: timm - Formats: Safetensors - Downloads in the last 30 days: 0 - Likes: 10 - License: apache-2.0 - Based on: KaraKaraWitch/Haruka-EVA02-base-448 - Fine-tunes: 1 - Tags: timm, onnx, image-classification, anime, danbooru, tagger, eva02, safetensors ## Links - Model: https://huggingface.co/KaraKaraWitch/haruka-tagger - Files: https://huggingface.co/KaraKaraWitch/haruka-tagger/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download KaraKaraWitch/haruka-tagger ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [KaraKaraWitch/haruka-tagger on Hugging Face](https://huggingface.co/KaraKaraWitch/haruka-tagger) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Haruka-Tagger (this page) | 0 | 10 | — | apache-2.0 | 2026-11-07 | | [MageTrail](https://mindket.com/p/a3c35252-cc66-4b2c-ab54-da01beda821a) | 134 | 49 | — | mit | 2026-10-14 | | [pixai-tagger-v1.0](https://mindket.com/p/9e6c2614-4591-4426-bb36-813dd434670d) | 36,625 | 165 | 486M | apache-2.0 | 2026-09-22 | | [PP-OCRv6\_manga](https://mindket.com/p/1ffe1df9-703d-4ee9-9cd8-e9bf6b756171) | 437 | 14 | — | apache-2.0 | 2026-11-06 | | [Gender-Classifier-Mini](https://mindket.com/p/75cd02bd-1729-43b4-b817-57d597517083) | 6,810,513 | 20 | 92.9M | apache-2.0 | 2025-03-28 | | [AIRealNet-2](https://mindket.com/p/180ca4e4-40b3-4670-b74b-7e4a958d9391) | 492 | 10 | 195M | mit | 2026-10-29 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-11-07 - First published: 2026-11-07 - Checked on Hugging Face: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open KaraKaraWitch/haruka-tagger on Hugging Face](https://huggingface.co/KaraKaraWitch/haruka-tagger) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/2eaa30c0-2cd7-4221-896b-0974b1653893) - [Its page on Mindket](https://mindket.com/p/2eaa30c0-2cd7-4221-896b-0974b1653893)