: HALO & Synthetic Haze Datasets

HALO

HALO covers 4 flare types across 32 distinct scenes, with multiple cameras and lighting variants per scene. Each sample carries three aligned 4K RGBA PNGs — the clean scene (gt), the same scene with flare added (flare), and the flare layer rendered against transparent background (separate) so it can be re-composited over arbitrary clean images at training time.

Per-scene composition

4,945 samples are stratified across 32 scenes; the four flare types share the same per-scene distribution at a 2 : 2 : 1 : 1 (Streak : Reflective : Glare : Shimmer) ratio.

Effect-type breakdown

Effect type	Samples	Share	Description
Streak	1,656	33.5%	bright streak / anamorphic stretch flares
Reflective	1,655	33.5%	inter-element ghost reflections
Glare	817	16.5%	wide soft halo + bloom
Shimmer	817	16.5%	iridescent / dispersive flare
Total	4,945	100%	32 scenes × 4 effects

Per-sample shard layout

halo/<base_name>.gt.png         clean scene without flare       (RGBA, 3840×2160)
halo/<base_name>.flare.png      same scene with flare added     (RGBA, 3840×2160)
halo/<base_name>.separate.png   flare-only on transparent bg    (RGBA, 3840×2160)
halo/<base_name>.json           per-sample metadata

Stream from Hugging Face

pip install -U "huggingface_hub[hf_xet]" webdataset pillow
hf auth login   # accept 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):
    meta = json.loads(s["json"])
    return {
        "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)
for sample in pipeline:
    print(sample["scene"], sample["effect"])
    break

Synthetic Haze

Per the paper's §7 atmospheric model, every clean image is augmented with multiple haze variants whose optical thickness τ follows depth predictions from Marigold, optionally modulated by Perlin noise for non-uniform haze and valley fog. Variants span homogeneous, non-homogeneous, indoor, outdoor, daytime, dense, and smoke-like conditions.

Source composition

Six upstream sources, dominated by HazeSpace2M (~82% of unique GTs). Each GT carries ~19–37 synthesized haze variants, for ~2 million total pairs.

Per-sample shard layout

<source>/<base_name>.gt.<ext>          clean RGB image (absent for OTS)
<source>/<base_name>.haze_NNN.png      synthesized haze variant
<source>/<base_name>.haze_NNN.txt      descriptive tag (e.g. out_fog_120)
<source>/<base_name>.json              per-sample metadata

<source> is one of hazespace, its, ots, wsrd, real_flare, istd. RESIDE-OTS clean GTs carry third-party photographer copyrights and are not bundled; the companion repo ships a fetch script that pulls the originals from the official RESIDE source.

Stream from Hugging Face

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

REPO = "jdzhang0929/uniser-haze-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):
    meta = json.loads(s["json"])
    haze_keys = sorted(k for k in s if k.startswith("haze_") and k.endswith(".png"))
    chosen = random.choice(haze_keys)
    gt_key = next((k for k in s if k.startswith("gt.")), None)
    return {
        "source": meta["source"],
        "gt":     Image.open(io.BytesIO(s[gt_key])).convert("RGB") if gt_key else None,
        "haze":   Image.open(io.BytesIO(s[chosen])).convert("RGB"),
        "tag":    s[chosen.replace(".png", ".txt")].decode(),
    }

pipeline = wds.WebDataset(urls, shardshuffle=True).shuffle(1000).map(decode)
for sample in pipeline:
    print(sample["source"], sample["tag"])
    break