Dataset Methods Results Gallery Human Eval Download

Synthetic Pipelines,
Controlled Condition Generation

Original IP2P Diffusion Rain
Original Rain IP2P
InstructPix2Pix + Physics Overlay

Snow and Rain Conditions

Text-guided diffusion editing (rain / snow_light / snow_heavy) with physics-based particle overlay. SSIM+LPIPS quality filtering.

24,356 rain (light + heavy)
49,991 snow (light + heavy)
2 intensities each
Original Day Night
Day Night
CycleGAN-Turbo (img2img)

Night Conditions

One-step image translation preserving scene structure. 1:1 mapping with FP16 precision on A100 GPUs.

23,974 night rows (3 subs)
+6,008 compound rain/snow_night
FP16 · Lanczos upscale
Original Fog Heavy
Original Fog
Koschmieder + Depth Anything V2

Fog Conditions

Physics-based atmospheric scattering with monocular depth from Depth Anything V2. Visibility sampled per row; intensity zone fixed by shard.

10,504 fog rows (3 subs)
3 zones (light/medium/heavy)
per-row visibility
Original Outpainted Small
Original Outpainted
FLUX.1-Fill-dev Outpainting

Scale / Distance Conditions

Source crop randomly downscaled and pasted onto a larger canvas. Bounding boxes re-mapped via per-image scale and paste offsets (deterministic affine).

7,323 small rows (3 subs)
2,823 / 3,000 / 1,500 cs10k / soda_voc / ktsh
bbox re-mapped per-row

End-to-End Pipeline

cs10k
44,105 rows • 3 classes • 14 GB
+
soda_voc
56,021 rows • 15 classes • 31 GB
+
soda_ktsh
22,030 rows • captions • 6.4 GB
↓

Weather Diffusion

IP2P + Physics

74,347

rain/snow light + heavy

Fog

Koschmieder + Depth Anything V2

10,504

light / medium / heavy

Day → Night

CycleGAN-Turbo

23,974

FP16 · Lanczos upscale

Night × Weather

Two-stage IP2P → CycleGAN

6,008

rain_night + snow_night (cs10k)

Outpainting

FLUX.1-Fill-dev

7,323

scale / distance

↓
Per-row quality_scores
DINOv3 · SSIM · CLIP · LPIPS
→
quality_alert flag
DINO < 0.85
→
Weather Classifier
SigLIP2 validation
→
FID / Texture
vs ACDC + Dempster-Shafer
↓
ConSynth-X Dataset
34,199 rows • 11 Parquet files (1 per condition) • 11 synthetic conditions • 16 unique object classes • multi-label annotations