LAION-Beyond: Reproducible Vision-Language Models Meet Concepts Out of Pre-Training

📄 Paper (CVPR 2025) | 💻 Code | 🌐 Project Page

Dataset Summary

LAION-Beyond is the first multi-domain benchmark specifically designed to evaluate the Out-of-Pre-training (OOP) generalization of vision-language models (e.g., CLIP, OpenCLIP, EVA-CLIP).

We distinguish two types of visual concepts:

• IP (In-Pre-training): concepts that appear in the pre-training data (e.g., LAION-400M / 2B / 5B)
• OOP (Out-of-Pre-training): concepts entirely absent from the pre-training data

Figure 1: Comparison between IP and OOP generalization. The former evaluates generalization within seen visual concepts, while the latter tests concepts absent during pre-training.

The key finding of our paper is that despite OpenCLIP's image encoder forming well-separated clusters for OOP concepts, zero-shot transfer fails significantly due to poor image-text alignment — the token embeddings for OOP class names were never aligned with visual features during pre-training.

---

Dataset Statistics

Split	Images	Concepts
OOP	106,052	674
IP	51,330	324
Total	157,382	998

Figure 2: (Left) Statistics of OOP/IP concepts across different LAION scales; (Right) Detailed train/val/test split in LAION-Beyond (400M).

Domains Covered:

• 🐾 Animals | 🏛️ Architecture | 👘 Attire
• 🎨 FolkArt | 🍜 Food | 🦋 Insects & Spiders
• 🗺️ Landmark | 🌿 Plants & Fungi | 🎮 Pokemon

Each domain contains an IP subset and an OOP subset, covering LAION-400M, LAION-2B, and LAION-5B scales to support neural scaling law research.

---

Dataset Structure

Each domain folder is named {Domain}{NumClasses}_{IP/OOP}, e.g., Animals42_IP, Animals92_OOP.

LAION_Beyond/
├── Animals42_IP/
│   ├── images/                        # jpg images organized by class
│   ├── label2name.json                # label index → class name
│   ├── name2label.json                # class name → label index
│   ├── merged_mapping.json            # merged label mapping
│   └── split_Xin_Animals42_IP.json    # train/val/test split info
├── Animals92_OOP/
│   └── ...
├── Architecture23_IP/
├── Architecture50_OOP/
├── Attire28_IP/
├── Attire54_OOP/
├── FolkArt27_IP/
├── FolkArt59_OOP/
├── Food27_IP/
├── Food53_OOP/
├── Insects_Spiders52_IP/
├── Insects_Spiders106_OOP/
├── Landmark30_IP/
├── Landmark59_OOP/
├── Plants_Fugi56_IP/
├── Plants_Fugi113_OOP/
├── Pokemon39_IP/
└── Pokemon89_OOP/

File Descriptions

File	Description
images/	Raw image files (JPG), organized by class subfolder
label2name.json	Mapping from integer label to class name string
name2label.json	Mapping from class name string to integer label
merged_mapping.json	Combined label mapping across splits
split_Xin_*.json	Train / val / test split assignments per image

---

Loading the Dataset

Option 1: Download full dataset (recommended)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="MHuangX/LAION-Beyond",
    repo_type="dataset",
    local_dir="./LAION_Beyond"
)

Option 2: Download a single domain only

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="MHuangX/LAION-Beyond",
    repo_type="dataset",
    local_dir="./LAION_Beyond",
    allow_patterns="Animals42_IP/**" 
)

---

Key Findings

1. Strong image features for OOP concepts: OpenCLIP's image encoder forms well-separated clusters for OOP concepts (clustering accuracy gap < 3% on most domains vs. IP concepts).

2. Image-text alignment failure: Zero-shot accuracy on OOP concepts is significantly lower than IP concepts, persisting even as pre-training data scales from 400M to 5B.

3. Name-tuning is the key: Our proposed FSNL and ZSNL algorithms, which fine-tune only the name (token) embeddings of OOP concepts, efficiently restore OOP generalization without degrading IP performance.

---

Algorithms

FSNL — Few-Shot Name Learning

Optimizes only OOP concept name embeddings using a few image-text pairs, with context augmentation via similar concept shuffling. Achieves state-of-the-art on 8/9 domains.

ZSNL — Zero-Shot Name Learning

Requires no image-text pairs. Uses Novel Class Discovery (NCD) and image-text bipartite graph matching to optimize OOP name embeddings from unlabeled images only.

---

Benchmark Results (400M split)

OOP Few-Shot Learning (4-shot, H-mean of OOP & IP accuracy)

Method	Animals	Architecture	Attire	FolkArt	Food	Insects	Landmark	Plants	Pokemon	Avg
OpenCLIP	26.75	30.75	25.88	35.04	15.36	22.38	40.25	21.43	24.48	26.92
CoOp	31.37	57.8	50.39	52.06	42.55	25.73	85.89	24.78	35.52	45.12
CLIP-Adapter	38.98	59.27	64.56	56.32	64.32	32.51	90.82	31.97	54.99	54.86
FSNL (ours)	46.17	62.63	71.65	63.03	70.0	44.03	94.48	44.12	68.87	62.55

---

Citation

If you use LAION-Beyond in your research, please cite:

@inproceedings{chen2025reproducible,
  title={Reproducible vision-language models meet concepts out of pre-training},
  author={Chen, Ziliang and Huang, Xin and Fan, Xiaoxuan and Wang, Keze and Zhou, Yuyu and Guan, Quanlong and Lin, Liang},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={14701--14711},
  year={2025}
}

---

License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

---

Authors

Xin Huang†, Ziliang Chen†, Xiaoxuan Fan, Keze Wang, Yuyu Zhou, Quanlong Guan, Liang Lin*

Affiliations: Peng Cheng Laboratory, Sun Yat-sen University, EPFL, Jinan University

†Equal Contribution · *Corresponding Author