FFHQ-2048 — NanoPocket Enhanced

The first new high-quality public face dataset since 2019. 1,000 sharp, artifact-free 2048×2048 portraits, derived from FFHQ and re-mastered with the NanoPocket Face Enhance model.

👉 Open the live drag-to-compare slider with mouse-wheel zoom

---

Why this dataset exists

FFHQ (NVIDIA, 2019) has been the gold-standard face dataset for the past five years — but the field has moved on. Modern generators (Flux, SD3 / SDXL, StyleGAN-T, portrait restoration nets) train at 1024² and above, and they expose every soft pixel, every JPEG ghost, every out-of-focus eyelash that the original FFHQ contains. Yet no comparable public face dataset has been released since FFHQ.

We built FFHQ-2048 to fill that gap:

• 2× spatial resolution — 1024×1024 → 2048×2048.
• Sharper, more detailed faces — pores, hair strands, iris texture, fabric weave.
• Artifact-free — no over-sharpening halos, no plastic skin, no identity drift.
• Filename-compatible with original FFHQ — 00000.png here corresponds to FFHQ index 0, so you can swap it in to existing pipelines without rewriting code.

This GitHub repo announces a public preview of the first 1,000 images. The full 70,000-image enhanced set is available on request — see Want the full set? below.

---

Before / after — six samples

GitHub strips iframes, so the table below shows static thumbnails. For an interactive drag-to-compare slider with mouse-wheel zoom, open the demo on Hugging Face Spaces ↗.

FFHQ index	Original — 1024²	Enhanced — 2048²
00046
00166
00394
00874
00894
00991

The differences look subtle at thumbnail scale. Open the interactive demo ↗ and zoom in — that's where the recovered detail (skin pores, iris texture, individual hair strands, fabric weave) becomes obvious.

---

Dataset summary

Field	Value
Number of images	1,000
Resolution	2048 × 2048
Format	PNG, lossless
Total size	~5.4 GB
Filename pattern	data/{index:05d}.png (e.g. data/00042.png)
Index range	00000 – 00999 (matches original FFHQ indices)
Source	NVIDIA FFHQ images1024x1024 first 1,000
Enhancement	NanoPocket Face Enhance

Nanopocket-ai/FFHQ-2048   (Hugging Face dataset)
├── README.md
└── data/
    ├── metadata.csv          # file_name, ffhq_index, original_split
    ├── 00000.png
    ├── 00001.png
    └── ... 998 more

data/metadata.csv lives next to the images, so the standard datasets ImageFolder loader picks it up automatically.

---

Quick start

pip install -U datasets huggingface_hub pillow

A runnable script with all three patterns lives in quickstart.py.

Option 1 — datasets library (with metadata)

from datasets import load_dataset

ds = load_dataset("Nanopocket-ai/FFHQ-2048", split="train")
print(ds)                       # 1000 rows: image, ffhq_index, original_split
print(ds[0]["image"].size)      # (2048, 2048)
print(ds[0]["ffhq_index"])      # 0
ds[0]["image"].save("sample.png")

Option 2 — huggingface_hub snapshot (full local copy)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="Nanopocket-ai/FFHQ-2048",
    repo_type="dataset",
    allow_patterns=["data/*", "README.md"],
)
print(local_dir)                # contains data/00000.png ... data/00999.png + data/metadata.csv

Option 3 — Single image via raw URL

import io
import requests
from PIL import Image

url = "https://huggingface.co/datasets/Nanopocket-ai/FFHQ-2048/resolve/main/data/00042.png"
img = Image.open(io.BytesIO(requests.get(url).content))
img.show()

---

Want the full enhanced FFHQ?

This repo is a public preview. We have re-mastered the entire 70,000-image FFHQ to 2048² with the same pipeline. If the previewed quality fits your research or product, get in touch:

Email: marketing@nanopocket.ai

Tell us briefly:

1. Who you are (lab / company / individual).
2. What you plan to use the data for.
3. Whether the use is non-commercial (FFHQ's CC BY-NC-SA 4.0 inheritance applies).

We will respond with a delivery method (LFS bundle, S3 link, or a private HF dataset invite).

---

About NanoPocket Face Enhance

NanoPocket Face Enhance is our in-house face restoration / super-resolution model, optimised to (a) preserve identity, (b) recover micro-detail (skin pores, hair, iris, lip texture) and (c) avoid the typical pitfalls of SR networks — over-sharpened halos, plastic skin, and waxy artifacts.

If you would like to use this model locally, please also contact: marketing@nanopocket.ai.

---

License & attribution

This dataset is a derivative work of NVIDIA's Flickr-Faces-HQ (FFHQ) dataset. Per FFHQ's license terms we inherit the same license and clearly indicate the changes we made.

• Dataset license: Creative Commons BY-NC-SA 4.0 — free use, redistribution and adaptation for non-commercial purposes, with attribution and share-alike.
• Per-image licenses: the underlying photographs were originally collected from Flickr under one of:
  • CC BY 2.0
  • CC BY-NC 2.0
  • Public Domain Mark 1.0
  • Public Domain CC0 1.0
  • U.S. Government Works Per-image author and license are recorded in the original ffhq-dataset-v2.json metadata, indexed by the same numeric IDs used in this repo.
• Indicated changes: every image has been upscaled 2× (1024×1024 → 2048×2048) and detail-enhanced by NanoPocket Face Enhance. No re-cropping, re-alignment or content edits beyond enhancement were performed.

Important — Not for facial recognition

Reproducing NVIDIA's explicit clause: this dataset is not intended for, and should not be used for, the development or improvement of facial recognition technologies.

---

Privacy & removal requests

We respect the same privacy / opt-out process as the upstream FFHQ dataset. To request removal of a photo of yourself:

1. On Flickr, do one of: tag the photo with no_cv, change the licence to All Rights Reserved or any CC NoDerivs variant, set the photo to private, or delete it.
2. Email researchinquiries@nvidia.com (the upstream maintainers) with your Flickr username.
3. Also email us at marketing@nanopocket.ai so we can remove the corresponding enhanced image from this repo and any future releases.

---

Citation

If you use this dataset, please cite both the original FFHQ paper and this release:

@inproceedings{karras2019stylebased,
  title     = {A Style-Based Generator Architecture for Generative Adversarial Networks},
  author    = {Tero Karras and Samuli Laine and Timo Aila},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2019},
  url       = {https://arxiv.org/abs/1812.04948}
}

@misc{nanopocket2026ffhq2048,
  title        = {FFHQ-2048: A 2K Re-master of Flickr-Faces-HQ via NanoPocket Face Enhance},
  author       = {NanoPocket},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Nanopocket-ai/FFHQ-2048}},
  note         = {Public preview release. Full 70k set available on request.}
}

---

Acknowledgements

• NVIDIA / NVlabs for releasing the original FFHQ dataset.
• Tero Karras, Samuli Laine, Timo Aila for the StyleGAN paper that introduced FFHQ.
• Vahid Kazemi & Josephine Sullivan for the face-alignment work that made the original collection possible.
• The Flickr photographers who shared their work under permissive licences.

---

Star ⭐ this repo

If FFHQ-2048 is useful to your work, starring this repo helps other researchers discover it and signals to us that the community wants the full 70k release prioritised.

---

Changelog

• v1.0 (2026-04) — Initial public release: first 1,000 images, 2048×2048, NanoPocket Face Enhance v1.