SynthBench

Benchmark dataset for evaluating synthetic data generation methods for visual classification. Contains real iPhone photos, FLUX text-to-image generated images, and programmatically augmented synthetic images across 6 object classes.

Classes

mouse, pen, phone, laptop, water bottle, Rubik's cube

Dataset Structure

data/
├── raw/                    # 308 raw iPhone photos (HEIC converted to JPEG)
├── real/                   # 309 preprocessed images (224x224), train/test split
│   ├── train/
│   └── test/
├── synthetic_t2i/          # 1,774 FLUX text-to-image generated images
├── synthetic_aug/          # 1,801 programmatically augmented synthetic images
└── synthetic_t2i_lowdiv/   # 597 low-diversity T2I images (ablation)

models/
├── real_baseline.pth       # ResNet-18 trained on real data only
├── t2i_zero_shot.pth       # ResNet-18 trained on T2I synthetic data (zero-shot)
└── aug_zero_shot.pth       # ResNet-18 trained on augmented synthetic data (zero-shot)

Usage

from huggingface_hub import snapshot_download

# Download everything
snapshot_download(repo_id="LakshC/SynthBench", repo_type="dataset")

# Download only real images
snapshot_download(repo_id="LakshC/SynthBench", repo_type="dataset", allow_patterns="data/real/*")

Or via CLI:

huggingface-cli download LakshC/SynthBench --repo-type dataset

Associated Code

GitHub: https://github.com/LakshC/SynthBench

Model

All checkpoints are ResNet-18 (via timm), fine-tuned for 6-class classification.