Artist Style Benchmark

A benchmark dataset of 34903 anime-style illustrations generated with AnimaImagine, each using a different Danbooru artist tag at fixed prompt/seed settings.

ASR Ranker Style Browser

This dataset also serves as the image host for ASR Ranker, an interactive browser for comparing and ranking artist styles.

• Browser: https://ranker.kuronet.top/browser
• Dataset: https://huggingface.co/datasets/Moeblack/artist-style-benchmark
• 选出最 hot 的画风吧。
• 总榜只有 trusted 用户才会加入（需要发送申请）；个人数据只针对自己本人。
• 所有数据来自我本人。
• 持续更新画师名中，目前有 23000+ 个画师名。

Dataset Structure

• image-files/ — sharded WebP thumbnails (768×987, quality 90), one per artist
• metadata.jsonl — one JSON object per image with:
  • image: repository-relative image path such as image-files/003/050926_42.webp
  • image_filename: image filename only, for display or local indexing
  • artist: Danbooru artist tag
  • danbooru_count: number of Danbooru posts for this artist
  • Full prompt tag fields (appearance_tags, outfit_tags, etc.)
  • Generation parameters (seed, steps, cfg_scale, aspect_ratio)
  • Images are sharded every 1000 source indices to avoid Hugging Face's 10,000-files-per-directory limit.

Do not infer the shard directory from the filename prefix. Use the image field from metadata.jsonl as the source of truth.

Image Direct Links

Use Hugging Face /resolve/ URLs for direct image access. The image field in metadata.jsonl is already a repository-relative path, so direct links can be constructed by appending it after /resolve/main/.

Example direct image URL:

https://huggingface.co/datasets/Moeblack/artist-style-benchmark/resolve/main/image-files/003/050926_42.webp

Markdown embedding example:

![sample](https://huggingface.co/datasets/Moeblack/artist-style-benchmark/resolve/main/image-files/003/050926_42.webp)

HTML embedding example:

<img src="https://huggingface.co/datasets/Moeblack/artist-style-benchmark/resolve/main/image-files/003/050926_42.webp" alt="sample">

Do not use /blob/ links for programmatic image loading. /blob/ opens the Hugging Face file viewer page and returns HTML; /resolve/ returns the actual WebP file.

Note about the legacy flat images/ directory

The old flat images/ directory from the initial partial upload has been removed. Current metadata and examples use only the sharded image-files/... layout, which avoids the Hub limit that rejects repositories with more than 10,000 files in a single directory.

Usage

import json
from huggingface_hub import hf_hub_download

# Load metadata
meta_path = hf_hub_download("Moeblack/artist-style-benchmark", "metadata.jsonl", repo_type="dataset")
with open(meta_path) as f:
    for line in f:
        row = json.loads(line)
        print(f"{row['artist']}: {row['danbooru_count']} posts")

# Get image URL (direct link)
url = f"https://huggingface.co/datasets/Moeblack/artist-style-benchmark/resolve/main/{row['image']}"

Generation Settings

All images generated with:

• Model: Anima base-v1.0
• Seed: 42 (fixed)
• Steps: 16
• CFG Scale: 7.0
• Aspect Ratio: 3:4 (~1MP, 896×1152)

License

MIT