InScene

InScene is a synthetic dataset of full-body / in-the-wild scene images, each containing a person, paired with rich metadata (generation prompt, face bounding box, identity ID, and 68-point facial landmarks). It was created for the paper Face2Scene: Using Facial Degradation as an Oracle for Diffusion-Based Scene Restoration (CVPR 2026), where facial degradation is used as a supervisory signal ("oracle") for diffusion-based restoration of full scenes.

Why this dataset

Faces are one of the most perceptually sensitive regions in an image and there exist strong, well-studied priors for facial quality. Face2Scene leverages this: by treating the face as an oracle for degradation, the model learns to restore the surrounding scene. To support this, InScene provides images in which a consistent set of identities appear across many diverse scenes, together with the localization metadata (face boxes and landmarks) needed to isolate and reason about the facial region. The val split additionally ships fixed low-quality (degraded) inputs paired with their clean targets, for consistent evaluation of restoration.

Data generation

Images were produced with the InfiniteYou-based data-generation pipeline available here:

• 🔧 Pipeline: https://github.com/amanwalia123/infiniteyoudatagen

The reference identity images used to condition generation are taken from the CelebRef-HQ dataset, available from the DMDNet repository:

• 🖼️ Reference images (CelebRef-HQ): https://github.com/csxmli2016/DMDNet

Please refer to the pipeline repository for the full generation details (identity conditioning, prompt construction, sampling settings, etc.).

Dataset structure

The dataset is split into train/ and val/. Each split is a flat directory of sample folders (one folder per generated image). The folder name encodes provenance: iden_<identity>_img_<source-face-id>_sample_<sample-id>_repeat_<k> (the same prompt/identity may be sampled multiple times).

train — each folder holds the generated high-quality image and its metadata:

train/iden_00001_img_11_sample_321_repeat_1/
  iden_00001_img_11_sample_321_repeat_1.png     # 1024x1024 RGB image (HQ)
  iden_00001_img_11_sample_321_repeat_1.json    # metadata

val — each folder additionally holds a degraded, low-quality version of the image prefixed LQ_. The un-prefixed *.png is the clean target and the LQ_*.png is the degraded restoration input:

val/iden_00003_img_4_sample_473_repeat_1/
  iden_00003_img_4_sample_473_repeat_1.png      # 1024x1024 RGB image (HQ target)
  LQ_iden_00003_img_4_sample_473_repeat_1.png   # degraded input
  iden_00003_img_4_sample_473_repeat_1.json     # metadata

Note: 6 of the val folders also contain a superseded copy of the trio prefixed old_ (old_*.png, old_LQ_*.png, old_*.json), kept for provenance.

Splits

Split	Sample folders	Identities	Files	Size
train	11,260	905	11,260 HQ PNG + 11,260 JSON	19 GB
val	1,329	100	1,329 HQ PNG + 1,329 LQ PNG + 1,329 JSON (+ 18 old_* leftover files)	3.4 GB

(The train and val identity sets are disjoint. train contains no LQ images — degradation for training is expected to be applied on the fly.)

Per-sample JSON schema

The JSON describes the HQ image (image_name is the un-prefixed PNG); for val, the degraded counterpart is the same name prefixed with LQ_.

Field	Type	Description
image_name	string	Filename of the paired HQ PNG.
prompt	string	Structured scene prompt used for generation (subject, framing, background, detail/style).
image_size	object	{ "width": 1024, "height": 1024 }.
face_bbox	list[int]	Face bounding box [x1, y1, x2, y2] in pixel coordinates.
identity	string	Zero-padded identity ID; the same ID denotes the same person across samples.
file_id	string	Source reference face-image ID (from CelebRef-HQ) used to condition this identity.
landmark	list[[int,int]]	68 facial landmark points [x, y] in pixel coordinates.

Usage

The dataset is a folder-of-folders (not a parquet/imagefolder layout), so the simplest way to use it is to download and glob the sample folders:

import glob, json, os
from PIL import Image
from huggingface_hub import snapshot_download

# Download one split (use allow_patterns="val/*" for val)
local = snapshot_download(
    "amir477/InScene",
    repo_type="dataset",
    allow_patterns="train/*",
)

samples = sorted(glob.glob(os.path.join(local, "train", "*")))
folder = samples[0]
name = os.path.basename(folder)

image = Image.open(os.path.join(folder, name + ".png"))      # 1024x1024 RGB (HQ)
meta = json.load(open(os.path.join(folder, name + ".json")))

print(image.size)            # (1024, 1024)
print(meta["identity"])      # e.g. "00001"
print(meta["face_bbox"])     # [x1, y1, x2, y2]
print(len(meta["landmark"])) # 68

For the val split, each sample also has a degraded input alongside the clean target:

val_folder = sorted(glob.glob(os.path.join(local_val, "val", "*")))[0]
name = os.path.basename(val_folder)
target = Image.open(os.path.join(val_folder, name + ".png"))         # clean HQ target
degraded = Image.open(os.path.join(val_folder, "LQ_" + name + ".png"))  # degraded input

Citation

If you find these images useful, please cite:

@inproceedings{kazerouni2026face2scene,
  title={Face2Scene: Using Facial Degradation as an Oracle for Diffusion-Based Scene Restoration},
  author={Kazerouni, Amirhossein and Suin, Maitreya and Aumentado-Armstrong, Tristan and Honari, Sina and Walia, Amanpreet and Mohomed, Iqbal and Derpanis, Konstantinos G and Taati, Babak and Levinshtein, Alex},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8428--8438},
  year={2026}
}