What it does

ACG2vec is a monorepo of deep-learning models and services built specifically for anime, comics, and games content. It covers text semantic search, image search by text or by image, image super-resolution, and even a rough “popularity score” predictor for illustrations. The author runs a live demo at cheerfun.dev/acg2vec.

The interesting bit

The project doesn’t just slap generic CLIP on anime and call it a day. The acgvoc2vec model was fine-tuned on 5.1 million sentence pairs scraped from Bangumi, pixiv, Moegirl, and Wikipedia—so it actually grasps that “Re:Zero” and “Rem” belong in the same semantic neighborhood. DCLIP further fine-tunes CLIP on the Danbooru2021 dataset with hand-rolled caption logic that prioritizes character and work tags over generic descriptors. There’s even a TensorFlow.js port of Real-CUGAN for browser-based super-resolution.

Key highlights

• acgvoc2vec: sentence-transformers fine-tuned on 5.1M ACG-specific sentence pairs (HuggingFace weights available)
• DCLIP: CLIP ViT-L/14 fine-tuned on Danbooru2021 with custom tag-to-caption preprocessing
• illust2vec: repurposes DeepDanbooru’s backbone as a feature extractor, then adds pooling to output 1024-dim image vectors
• pix2score: multi-task ResNet101 predicting binned bookmark counts, view counts, and “sanity level” (NSFW-ishness)
• real-cugan_tf: TensorFlow.js implementation of the Real-CUGAN super-resolution model, runnable in-browser
• Webapp module provides “out-of-the-box” services for tag prediction, image search, and feature extraction

Caveats

• Docker deployment module is explicitly noted as “未开发完成” (not finished)
• pix2score is marked as “训练中” (still training), and the README cuts off mid-sentence in the Pix2Score section—suggesting the docs are incomplete or the project is somewhat maintenance-stretched
• The architecture diagram and model descriptions are thorough, but you’ll be reading Chinese for the finer details

Verdict

Worth a look if you’re building search, recommendation, or tagging tools for ACG content and need domain-tuned embeddings without training them yourself. Skip it if you want a polished, production-ready platform—the live demo is nice, but the Docker setup is unfinished and several models are still works-in-progress.