NearID — Multi-View Identity Dataset

This is the base positives dataset for the NearID project. Each sample contains multiple views of the same identity rendered in different backgrounds/contexts.

Near-identity distractors (different but similar instances in matched context) are available as separate datasets listed below. Together, they form the NearID training and evaluation benchmark.

Quick Start

from datasets import load_dataset

# Load base positives
ds = load_dataset("Aleksandar/NearID")

# Load a negative source for contrastive training/evaluation
neg = load_dataset("Aleksandar/NearID-Flux")

Dataset Structure

Column	Type	Description
id	int64	Sample ID (shared across all NearID datasets). Sparse — not the row position
category	string	Object category
category_description	string	Natural language description of the identity
images1, images2, images3	image	Multi-view images of the same identity in different contexts
masks1, masks2, masks3	image	Foreground object masks for each view
n_images	int64	Number of valid views (2 or 3)
objaverse_id	string	Source Objaverse object identifier
prompts1–prompts3	string	Generation prompts for each view
filenames1–filenames3	string	Original source filenames
size1–size3	sequence[int64]	Pixel dimensions of each view
dino_01, dino_02, dino_12	float64	Pairwise DINO similarity between views (curation signal)
aesthetics1–aesthetics3	float64	Aesthetic score per view (curation signal)
quality	string	Quality label
id_safe	bool	Whether the identity passed the safety filter
split	string	Official split: train (18,786), val (100) or test (500)

Splits. The official train/val/test partition used for every result in the paper is available two equivalent ways: filter on the split column, or use the JSON index files in github.com/Gorluxor/NearID/tree/master/splits. Those files list 0-based row positions, not id values.

import json
from datasets import load_dataset

ds = load_dataset("Aleksandar/NearID", split="train")   # 19,386 rows — do not shuffle
test = ds.filter(lambda r: r["split"] == "test")        # 500 rows
# equivalently: ds.select(json.load(open("splits/test.json")))

All NearID Datasets

Dataset	Description	Resolution
Aleksandar/NearID	Multi-view positives (anchor + positive views)	Base
Aleksandar/NearID-Flux	Near-identity distractors via FLUX.1 inpainting	512×512
Aleksandar/NearID-Flux_1024	Near-identity distractors via FLUX.1 inpainting	1024×1024
Aleksandar/NearID-FluxC	Near-identity distractors via FLUX.1 Canny-guided inpainting	512×512
Aleksandar/NearID-FluxC_1024	Near-identity distractors via FLUX.1 Canny-guided inpainting	1024×1024
Aleksandar/NearID-PowerPaint	Near-identity distractors via PowerPaint inpainting	512×512
Aleksandar/NearID-Qwen	Near-identity distractors via Qwen-based inpainting	512×512
Aleksandar/NearID-Qwen_1328	Near-identity distractors via Qwen-based inpainting	1328×1328
Aleksandar/NearID-SDXL	Near-identity distractors via Stable Diffusion XL inpainting	512×512
Aleksandar/NearID-SDXL_1024	Near-identity distractors via Stable Diffusion XL inpainting	1024×1024

Related

• Model: Aleksandar/nearid-siglip2 — NearID identity embedding model
• Paper: NearID: Identity Representation Learning via Near-identity Distractors
• Code: github.com/Gorluxor/NearID

License & Attribution

This dataset is released under CC-BY-4.0. It is derived from the SynCD dataset (MIT License, Copyright 2022 SynCD). If you use this dataset, please cite both NearID and SynCD.

Citation

@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}
}