Downloads · 30 days
0
geonusaf-stage3-fakepairs-block-fold1
sugam24/geonusaf-stage3-fakepairs-block-fold1
geonusaf-stage3-fakepairs-block-fold1 is a machine learning model from sugam24. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-4.0.
Synthetic Kathmandu Valley satellite image / land-use mask pairs, produced by chaining the Part-2 Stage-1 mask diffusion model into the Stage-2 mask-conditioned image generator. Intended as training-side augmentation…
Access
Public
Updated Aug 31, 2026
Repo size
10.7 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.gz5.7 GB · 99%
At a glance
- License
- cc-by-4.0
- Access
- Public
- Created
- Aug 30, 2026
- Updated
- Aug 31, 2026
- SHA
- 8b917f00
Model card
From the Hugging Face model README
GeoNUSAF Stage 3 - synthetic image-mask pairs (block fold 1)
Synthetic Kathmandu Valley satellite image / land-use mask pairs, produced by chaining the Part-2 Stage-1 mask diffusion model into the Stage-2 mask-conditioned image generator. Intended as training-side augmentation for the Part-1 segmentation models.
| field | value |
|---|---|
| mask generator | sugam24/geonusaf-stage1-maskddpm-onehot-block-fold1 (epoch 1399) |
| image generator | Pranilllllll/geonusaf-stage2-mask2image-block-fold1-D (step 7000) |
| split | block fold 1, train_sha1 eb0aebdd9919 |
| real training pairs | 804 |
| sets | R1 = 804 pairs, R2 = 1608 pairs, R3 = 2412 pairs |
| draws | independent per set |
| mask sampling | DDIM 200, pool ratio 2.5, post-processing modal3/close0/thin30/bulk0 |
| image sampling | DDIM 30, guidance 3.5, cond scale 1.0, canny True @ 0.4 |
| resolution | 512x512, 1.374 m/px |
| scorer | Pranilllllll/segformer-b0-kaggle-ktm-dataset-block-fold1, ceiling mIoU 0.4646 |
Quality
| set | layout mIoU | ratio to ceiling | KID vs val | FID vs val | memorization ratio (feat / pixel / mask) |
|---|---|---|---|---|---|
| R1 | 0.4395 | 0.946 | 0.04698 +/- 0.00925 | 128.7 | 0.99 / 0.89 / 1.39 |
| R2 | 0.4375 | 0.942 | 0.05243 +/- 0.00936 | 132.2 | 1.00 / 0.90 / 1.34 |
| R3 | 0.4337 | 0.933 | 0.05179 +/- 0.00994 | 133.0 | 1.01 / 0.90 / 1.42 |
Memorization ratio = mean nearest-neighbour distance of a fake to the training set, divided by the same statistic for a held-out real tile. At or above 1.0 the fakes are no closer to the training data than a genuinely new real tile is.
Contract
genimage{i}.png- 512x512x3uint8, BGR on disk (written withcv2.imwrite, read withcv2.IMREAD_COLOR)genmask{i}.png- 512x512 single-channeluint8, values{0..5}semantic +255ignore- class order: 0=Residential, 1=Road, 2=River, 3=Forest, 4=UnusedLand, 5=Agricultural
Usage restriction
Training side only. These pairs must never enter a validation or test split in any fold. The generators saw block fold 1's TRAIN tiles; putting their output into an evaluation set is leakage.
- License
- cc-by-4.0
- Created
- Aug 30, 2026
- Updated
- Aug 31, 2026