zyyyz/ID-Bench

---
dataset_info:
features:
- name: target_image
dtype: image
- name: condition_image_1
dtype: image
- name: condition_image_2
dtype: image
- name: condition_image_3
dtype: image
- name: caption
dtype: string
splits:
- name: train
num_bytes: 5757742734
num_examples: 2000
download_size: 5736519068
dataset_size: 5757742734
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: cc-by-4.0
task_categories:
- image-to-image
language:
- zh
- en
size_categories:
- 10K<n<100K
---

# ID-Bench

**ID-Bench** is a real-world benchmark for **multi-reference identity-preserving image generation**. It is built from real-world e-commerce advertising images and organized by **product identity**, with the goal of evaluating whether a model can generate a **novel target image** that both preserves product identity and follows target-specific variation cues.

We release, in this repository, the curated benchmark dataset that is consistent with the one used for evaluation in our paper. The full training dataset will be released soon.

## What Makes ID-Bench Different?

Recent progress in reference-guided image generation has improved visual quality substantially, but evaluating **multi-reference identity-preserving generation** remains challenging. Many existing settings do not clearly distinguish:

- genuine same-identity generation,
- trivial copying from the conditioning set,
- and target-guided variation control.

ID-Bench is designed to address this gap in a real-world product-image setting.

## Data Format

Each example contains:

- `target_image`: the held-out target image
- `condition_image_1`: reference image 1 from the same product identity
- `condition_image_2`: reference image 2 from the same product identity
- `condition_image_3`: reference image 3 from the same product identity
- `caption`: a short target-oriented caption describing the desired image or scene

For more information, please visit our project website: [ID-Bench](https://zyyyz.github.io/IDBench/).