Aleksandar/NearID-SDXL

---
language:
- en
license: cc-by-4.0
size_categories:
- 10K<n<100K
task_categories:
- image-feature-extraction
- image-to-image
pretty_name: NearID-SDXL (Near-Identity Distractors)
dataset_info:
features:
- name: id
dtype: int64
- name: category
dtype: string
- name: category_description
dtype: string
- name: nimg1
dtype: image
- name: nimg2
dtype: image
- name: nimg3
dtype: image
- name: n_images
dtype: int64
- name: objaverse_id
dtype: string
- name: prompts1
dtype: string
- name: prompts2
dtype: string
- name: prompts3
dtype: string
- name: quality
dtype: string
splits:
- name: train
tags:
- nearid
- near-identity-distractors
- identity-embedding
- inpainting
- synthetic
- metric-learning
---

# NearID-SDXL — Near-Identity Distractors (Stable Diffusion XL inpainting)

[](https://huggingface.co/Aleksandar/nearid-siglip2) [](https://huggingface.co/papers/2604.01973) [](https://gorluxor.github.io/NearID/) [](https://github.com/Gorluxor/NearID) [](https://www.kaust.edu.sa/) [](https://research.snap.com/)

This dataset contains **near-identity distractors** generated by **Stable Diffusion XL inpainting** at **512×512** resolution as part of the [NearID](https://huggingface.co/Aleksandar/nearid-siglip2) project.

Each sample contains up to 3 distractor images (`nimg1`, `nimg2`, `nimg3`): different but visually similar instances inpainted into the **exact same background/context** as the corresponding anchor in the base [Aleksandar/NearID](https://huggingface.co/datasets/Aleksandar/NearID) dataset. These distractors are used to train and evaluate identity embeddings that distinguish true identity from contextual shortcuts.

## Quick Start

```python
from datasets import load_dataset

# Load this negative source
ds = load_dataset("Aleksandar/NearID-SDXL")

# Load base positives for anchor/positive pairs
positives = load_dataset("Aleksandar/NearID")
```

## Dataset Structure

| Column | Type | Description |
|---|---|---|
| `id` | int64 | Sample ID (matches the base NearID dataset) |
| `category` | string | Object category (`rigid`) |
| `category_description` | string | Natural language description of the object |
| `nimg1`, `nimg2`, `nimg3` | image | Near-identity distractor images (up to 3 per sample) |
| `n_images` | int64 | Number of valid distractor images |
| `objaverse_id` | string | Source Objaverse object identifier |
| `prompts1`, `prompts2`, `prompts3` | string | Generation prompts for each distractor |
| `quality` | string | Quality label |

## How the Distractors Were Generated

1. For each anchor identity in the base NearID dataset, a semantically similar but **different** object instance was retrieved.
2. The distractor instance was inpainted into the **same background** as the anchor using **Stable Diffusion XL inpainting**.
3. Resolution: **512×512** pixels.

This creates a controlled test: a model must rely on intrinsic identity features, not background context, to distinguish anchor from distractor.

## All NearID Datasets

| Dataset | Description | Resolution |
|---|---|---|
| [Aleksandar/NearID](https://huggingface.co/datasets/Aleksandar/NearID) | Multi-view positives (anchor + positive views) | Base |
| [Aleksandar/NearID-Flux](https://huggingface.co/datasets/Aleksandar/NearID-Flux) | Near-identity distractors via FLUX.1 inpainting | 512×512 |
| [Aleksandar/NearID-Flux_1024](https://huggingface.co/datasets/Aleksandar/NearID-Flux_1024) | Near-identity distractors via FLUX.1 inpainting | 1024×1024 |
| [Aleksandar/NearID-FluxC](https://huggingface.co/datasets/Aleksandar/NearID-FluxC) | Near-identity distractors via FLUX.1 Canny-guided inpainting | 512×512 |
| [Aleksandar/NearID-FluxC_1024](https://huggingface.co/datasets/Aleksandar/NearID-FluxC_1024) | Near-identity distractors via FLUX.1 Canny-guided inpainting | 1024×1024 |
| [Aleksandar/NearID-PowerPaint](https://huggingface.co/datasets/Aleksandar/NearID-PowerPaint) | Near-identity distractors via PowerPaint inpainting | 512×512 |
| [Aleksandar/NearID-Qwen](https://huggingface.co/datasets/Aleksandar/NearID-Qwen) | Near-identity distractors via Qwen-based inpainting | 512×512 |
| [Aleksandar/NearID-Qwen_1328](https://huggingface.co/datasets/Aleksandar/NearID-Qwen_1328) | Near-identity distractors via Qwen-based inpainting | 1328×1328 |
| [Aleksandar/NearID-SDXL](https://huggingface.co/datasets/Aleksandar/NearID-SDXL) | Near-identity distractors via Stable Diffusion XL inpainting | 512×512 | **← this dataset** |
| [Aleksandar/NearID-SDXL_1024](https://huggingface.co/datasets/Aleksandar/NearID-SDXL_1024) | Near-identity distractors via Stable Diffusion XL inpainting | 1024×1024 |

## Related

- **Model:** [Aleksandar/nearid-siglip2](https://huggingface.co/Aleksandar/nearid-siglip2) — NearID identity embedding model
- **Paper:** [NearID: Identity Representation Learning via Near-identity Distractors](https://huggingface.co/papers/2604.01973)
- **Code:** [github.com/Gorluxor/NearID](https://github.com/Gorluxor/NearID)

## License & Attribution

This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). It is derived from the [SynCD](https://github.com/nupurkmr9/syncd) dataset (MIT License, Copyright 2022 SynCD). If you use this dataset, please cite both NearID and SynCD.

## Citation

```bibtex
@article{cvejic2026nearid,
title={NearID: Identity Representation Learning via Near-identity Distractors},
author={Cvejic, Aleksandar and Abdal, Rameen and Eldesokey, Abdelrahman and Ghanem, Bernard and Wonka, Peter},
journal={arXiv preprint arXiv:2604.01973},
year={2026}
}
```

language:
- 英语
license: 知识共享署名4.0（CC-BY-4.0）
size_categories:
- 10000 < 样本数量 < 100000
task_categories:
- 图像特征提取
- 图像到图像转换
pretty_name: NearID-SDXL（近身份干扰项）
dataset_info:
features:
- name: 样本ID
dtype: int64
- name: 类别
dtype: 字符串
- name: 类别描述
dtype: 字符串
- name: nimg1
dtype: 图像
- name: nimg2
dtype: 图像
- name: nimg3
dtype: 图像
- name: 有效图像数量
dtype: int64
- name: Objaverse ID
dtype: 字符串
- name: 提示词1
dtype: 字符串
- name: 提示词2
dtype: 字符串
- name: 提示词3
dtype: 字符串
- name: 质量标签
dtype: 字符串
splits:
- name: 训练集
tags:
- nearid
- 近身份干扰项（near-identity distractors）
- 身份嵌入（identity embedding）
- 图像修复（inpainting）
- 合成数据集（synthetic）
- 度量学习（metric-learning）

# NearID-SDXL — 近身份干扰项（Stable Diffusion XL 图像修复）

[](https://huggingface.co/Aleksandar/nearid-siglip2) [](https://huggingface.co/papers/2604.01973) [](https://gorluxor.github.io/NearID/) [](https://github.com/Gorluxor/NearID) [](https://www.kaust.edu.sa/) [](https://research.snap.com/)

本数据集属于[NearID](https://huggingface.co/Aleksandar/nearid-siglip2)项目的一部分，包含由**Stable Diffusion XL 图像修复（inpainting）**生成的**近身份干扰项（near-identity distractors）**，分辨率为**512×512**像素。

每个样本最多包含3张干扰项图像（`nimg1`、`nimg2`、`nimg3`）：这些图像与基础数据集[Aleksandar/NearID](https://huggingface.co/datasets/Aleksandar/NearID)中的对应锚点图像视觉相似但内容不同，且被修复至与锚点完全一致的背景/场景中。此类干扰项可用于训练和评估能够从上下文捷径中区分真实身份的身份嵌入模型。

## 快速入门

python
from datasets import load_dataset

# 加载本负样本数据集
ds = load_dataset("Aleksandar/NearID-SDXL")

# 加载锚点/正样本对的基础正样本集
positives = load_dataset("Aleksandar/NearID")

## 数据集结构

| 列名 | 数据类型 | 描述 |
|---|---|---|
| `id` | int64 | 样本ID（与基础NearID数据集一致） |
| `category` | 字符串 | 对象类别（`rigid`，即刚性物体） |
| `category_description` | 字符串 | 对象的自然语言描述 |
| `nimg1`、`nimg2`、`nimg3` | 图像 | 近身份干扰项图像（每个样本最多3张） |
| `n_images` | int64 | 有效干扰图像的数量 |
| `objaverse_id` | 字符串 | 来源Objaverse对象标识符 |
| `prompts1`、`prompts2`、`prompts3` | 字符串 | 每张干扰项的生成提示词 |
| `quality` | 字符串 | 质量标签 |

## 干扰项生成流程

1. 针对基础NearID数据集中的每个锚点身份，检索得到一个语义相似但内容不同的对象实例。
2. 使用**Stable Diffusion XL 图像修复（inpainting）**将该干扰实例修复至与锚点完全一致的背景中。
3. 分辨率：**512×512**像素。

本数据集构建了可控测试场景：模型必须依赖内在的身份特征而非背景上下文，才能区分锚点与干扰项。

## 全系列NearID数据集

| 数据集 | 描述 | 分辨率 |
|---|---|---|
| [Aleksandar/NearID](https://huggingface.co/datasets/Aleksandar/NearID) | 多视角正样本（锚点+正样本视角） | 基础分辨率 |
| [Aleksandar/NearID-Flux](https://huggingface.co/datasets/Aleksandar/NearID-Flux) | 基于FLUX.1图像修复生成的近身份干扰项 | 512×512 |
| [Aleksandar/NearID-Flux_1024](https://huggingface.co/datasets/Aleksandar/NearID-Flux_1024) | 基于FLUX.1图像修复生成的近身份干扰项 | 1024×1024 |
| [Aleksandar/NearID-FluxC](https://huggingface.co/datasets/Aleksandar/NearID-FluxC) | 基于FLUX.1 Canny引导图像修复生成的近身份干扰项 | 512×512 |
| [Aleksandar/NearID-FluxC_1024](https://huggingface.co/datasets/Aleksandar/NearID-FluxC_1024) | 基于FLUX.1 Canny引导图像修复生成的近身份干扰项 | 1024×1024 |
| [Aleksandar/NearID-PowerPaint](https://huggingface.co/datasets/Aleksandar/NearID-PowerPaint) | 基于PowerPaint图像修复生成的近身份干扰项 | 512×512 |
| [Aleksandar/NearID-Qwen](https://huggingface.co/datasets/Aleksandar/NearID-Qwen) | 基于Qwen的图像修复生成的近身份干扰项 | 512×512 |
| [Aleksandar/NearID-Qwen_1328](https://huggingface.co/datasets/Aleksandar/NearID-Qwen_1328) | 基于Qwen的图像修复生成的近身份干扰项 | 1328×1328 |
| [Aleksandar/NearID-SDXL](https://huggingface.co/datasets/Aleksandar/NearID-SDXL) | 基于Stable Diffusion XL图像修复生成的近身份干扰项 | 512×512 | **← 本数据集** |
| [Aleksandar/NearID-SDXL_1024](https://huggingface.co/datasets/Aleksandar/NearID-SDXL_1024) | 基于Stable Diffusion XL图像修复生成的近身份干扰项 | 1024×1024 |

## 相关资源

- **模型**：[Aleksandar/nearid-siglip2](https://huggingface.co/Aleksandar/nearid-siglip2) — NearID身份嵌入模型
- **论文**：[NearID: 基于近身份干扰项的身份表征学习](https://huggingface.co/papers/2604.01973)
- **代码**：[github.com/Gorluxor/NearID](https://github.com/Gorluxor/NearID)

## 许可与署名

本数据集采用[知识共享署名4.0（CC-BY-4.0）](https://creativecommons.org/licenses/by/4.0/)许可发布，其衍生自[SynCD](https://github.com/nupurkmr9/syncd)数据集（采用MIT许可，版权归2022年SynCD所有）。若使用本数据集，请同时引用NearID与SynCD。

## 引用格式

bibtex
@article{cvejic2026nearid,
title={NearID: Identity Representation Learning via Near-identity Distractors},
author={Cvejic, Aleksandar and Abdal, Rameen and Eldesokey, Abdelrahman and Ghanem, Bernard and Wonka, Peter},
journal={arXiv preprint arXiv:2604.01973},
year={2026}
}