This dataset is released for non-commercial academic research only. By requesting access you agree to: (1) use it solely for non-commercial research; (2) cite our CVPR 2026 paper (UniSER) when publishing results derived from this dataset; (3) not redistribute the dataset under terms more permissive than CC-BY-NC-SA-4.0; (4) attribute the public HDRI environments (Poly Haven / freepoly.org) used as scene backgrounds.

I agree to the above terms.

HALO Synthetic Lens-Flare Dataset ✨

HALO is a 3D-rendered synthetic lens-flare dataset companion to our CVPR 2026 paper, UniSER: A Foundation Model for Unified Soft Effects Removal. Each sample is a Blender-rendered 4K scene with a paired flare-free / flared / flare-only triplet — designed for training and benchmarking single-image lens-flare removal.

📦 Size: ~150 GB / 4945 samples / ~80 WebDataset shards 📄 Paper: arXiv 2511.14183 🐙 Code: github.com/Evergreen0929/UniSER-Datasets

📦 Dataset Composition

4945 4K-rendered samples spanning 32 scenes × 4 flare types.

Effect type	Count	Description
Streak	1,656	bright streak / anamorphic stretch flares
Reflective	1,655	inter-element ghost reflections
Glare	817	wide soft halo + bloom
Shimmer	817	iridescent / dispersive flare
Total	4,945

Scenes cover indoor + outdoor + day + night + dusk under varied HDRI lighting (sourced from Poly Haven and freepoly.org). All scenes are 3D-rendered with no real-world photography; each frame is shipped at 3840×2160 (4K) RGBA PNG.

Per-scene counts

Scene	Total	Streak	Reflective	Glare	Shimmer
Scene011	302	101	101	50	50
Scene075	254	84	86	42	42
Scene039	253	85	84	42	42
Scene022	252	84	84	42	42
Scene013	226	77	74	37	38
Scene025	204	68	69	34	33
Scene041	204	68	69	34	33
Scene017, 045	203	68–69	68–69	33	33
Scene021, 029, 033, 049, 051	202	67–69	67–68	33	33
Scene072	201	68	67	33	33
Scene005	152	50	52	25	25
Scene003, 007, 023, 037, 053	150	50	50	25	25
Scene048	118	39	39	20	20
Scene054	115	38	37	20	20
Scene071	101	33	34	17	17
Scene015, 064	100	33	33	17	17
Scene065	49	16	17	8	8
Scene077	48	17	16	7	8
Scene057, 069	34	11	11	6	6
Scene035, 061	16	6	6	2	2

🗂️ Format

Sharded in WebDataset format. Each shards/shard-NNNNN.tar contains samples whose tar member names follow:

halo/{base_name}.gt.png         clean scene without flare  (RGBA, 3840×2160)
halo/{base_name}.flare.png      same scene with flare      (RGBA, 3840×2160)
halo/{base_name}.separate.png   flare-only layer           (RGBA, 3840×2160)
halo/{base_name}.json           per-sample metadata

base_name encodes {scene}_{effect_id}_{camera} (e.g. Scene003_Glare001_camera01). WebDataset groups entries sharing the same prefix-before-the-first-dot into one sample.

The triplet supports three task formulations:

flare → light (paired flare removal)
light → flare (forward flare synthesis)
flare = light + separate (additive decomposition; separate is the flare layer alone)

Per-sample JSON metadata schema:

{
  "scene":       "Scene003",
  "effect_type": "Glare",
  "effect_id":   "Glare001",
  "sample_id":   "Scene003_Glare001_camera01",
  "orig_idx":    7313,
  "final_idx":   0,
  "light":       "Scene003_Glare001_camera01.gt.png",
  "flare":       "Scene003_Glare001_camera01.flare.png",
  "separate":    "Scene003_Glare001_camera01.separate.png"
}

🚀 Quick Start

pip install -U "huggingface_hub[hf_xet]" webdataset pillow
hf auth login
# Then visit and accept the terms at:
#   https://huggingface.co/datasets/jdzhang0929/halo-flare-dataset

import io, json
from huggingface_hub import HfFileSystem
import webdataset as wds
from PIL import Image

REPO = "jdzhang0929/halo-flare-dataset"
urls = [
    f"https://huggingface.co/datasets/{REPO}/resolve/main/{p[len(f'datasets/{REPO}/'):]}"
    for p in HfFileSystem().ls(f"datasets/{REPO}/shards", detail=False)
    if p.endswith(".tar")
]

def decode(s):
    if "json" not in s:
        return None
    meta = json.loads(s["json"])
    return {
        "sample_id":  meta["sample_id"],
        "scene":      meta["scene"],
        "effect":     meta["effect_id"],
        "light":      Image.open(io.BytesIO(s["gt.png"])).convert("RGB"),
        "flare":      Image.open(io.BytesIO(s["flare.png"])).convert("RGB"),
        "separate":   Image.open(io.BytesIO(s["separate.png"])),  # keep RGBA
    }

pipeline = (wds.WebDataset(urls, shardshuffle=True)
              .shuffle(500)
              .map(decode)
              .select(lambda x: x is not None))

for sample in pipeline:
    print(sample["sample_id"], sample["scene"], sample["effect"])
    break

finalist.json ships alongside the shards as a per-record reference table (record-to-shard lookup is via sample_id).

⚠️ Caveats

All scenes are 3D-rendered; no real-world photographs are included.
Image-Based-Lighting environments are sourced from external HDRI providers — see Acknowledgments below for full attribution.

🙏 Acknowledgments

The HALO renders use environment maps and skyboxes from public Image-Based-Lighting libraries. The 4945 samples in this release span 118 distinct HDRI environments sourced from the following providers; we are grateful to the authors and contributors for releasing them under permissive terms.

Poly Haven (primary HDRI source)

The majority of background environments are HDRIs from Poly Haven, a community-supported public-domain library of high-resolution HDRIs released under CC0 1.0 Universal (public domain, no attribution required, but acknowledged here per academic convention).

A non-exhaustive list of Poly Haven HDRIs used:

Category	Sample environments (Poly Haven asset slugs)
Coast / harbour	`simon-s-town-harbour`, `blouberg-sunrise-1/2`, `tucker-wreck`, `aristea-wreck`, `abandoned-slipway`, `abandoned-ships-2`, `blue-lagoon`, `seaside-hills-under-the-scorching-sun`
Mountain / valley	`kiara-4-mid-morning`, `kiara-9-dusk`, `rustig-koppie`, `goegap`, `mpumalanga-veld`, `birchwood`, `clarens-midday`, `clarens-night-01`, `aerial-canadian-mountains-17k`
Urban / heritage	`rhodes-memorial`, `bismarckturm-hillside`, `teufelsberg-roof`, `roman-s-ruins`, `abandoned-church`, `venice-dawn-1/2`, `venice-sunset`, `shanghai-riverside`, `hongkong-airport-north`
Industrial / agricultural	`versveldpas`, `abandoned-tank-farm-01/05`, `brick-factory-02`, `quarry-02`, `sunset-in-the-chalk-quarry`, `straw-rolls-field-01`, `veld-fire`, `path-by-farm-lands-sunny-day`
Pastoral / natural	`autumn-meadow`, `flower-field-day`, `green-lawn`, `winter-lake-01`, `blau-river`, `forst-lake-green-blue-sky`, `garden-{summer,fall,winter}-backdrop-day`, `szymanski-park`, `leisure-square`
Aerial landscape	`aerial-canadian-landscape`, `aerial-mountain-landscape`, `aerial-cityscape-twilight-sunset`, `aerial-view-of-the-plains`, `aerial-sunset-at-lake`, `aerial-suburban-city-sunny`, `aerial-dramatic-sunset`

The complete asset-to-record mapping is recoverable from the light_image_key field in finalist.json (each path's _4K/ parent directory encodes the source HDRI slug). HDRI slugs match the canonical names on polyhaven.com/hdris.

freepoly.org

A subset of renders (Scene029, 29 samples) use the housetop-05-freepoly-org HDRI sourced from freepoly.org. We thank the freepoly.org maintainers for hosting freely-available 3D resources.

Other thanks

Rendering pipeline built on Blender (GPL-3.0).
Dataset card structure and WebDataset packaging convention adapted from our companion release, UniSER-Datasets — see that repo for upstream haze dataset acknowledgments.

📜 License

This release is distributed under CC-BY-NC-SA-4.0 (non-commercial academic research only).

Upstream HDRI environments retain their original licenses, all of which are equal to or more permissive than CC-BY-NC-SA-4.0:

Poly Haven HDRIs: CC0 1.0 Universal (public domain)
freepoly.org HDRIs: see freepoly.org terms

📚 Citation

@article{zhang2025uniser,
  title={UniSER: A Foundation Model for Unified Soft Effects Removal},
  author={Zhang, Jingdong and Zhang, Lingzhi and Liu, Qing and Chiu, Mang Tik and Barnes, Connelly and Wang, Yizhou and You, Haoran and Liu, Xiaoyang and Zhou, Yuqian and Lin, Zhe and others},
  journal={arXiv preprint arXiv:2511.14183},
  year={2025}
}

📮 Contact

Please contact Jingdong Zhang with any questions about the HALO dataset.

Downloads last month: -

Total file size:

153 GB

Paper for jdzhang0929/halo-flare-dataset

UniSER: A Foundation Model for Unified Soft Effects Removal

Paper • 2511.14183 • Published Apr 28