Downloads · 30 days
0
sugam24/geonusaf-stage1-maskddpm-onehot-block-fold1
geonusaf-stage1-maskddpm-onehot-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. It is set up for diffusers.
Generates novel Kathmandu Valley land-use layouts as single-channel masks. Trained on the fold-1 TRAIN tiles only; validation tiles never enter a gradient.
Downloads · 30 days
0
Access
Public
Updated Aug 29, 2026
Repo size
30.3 GB
Likes
0
Public
Click a slice to open those files.
.pt372 MB · 99%
From the Hugging Face model README
Generates novel Kathmandu Valley land-use layouts as single-channel masks. Trained on the fold-1 TRAIN tiles only; validation tiles never enter a gradient.
Arm B of the label-representation ablation: one-hot channels. Arm A used the SatSynth binary code and is in a separate repo.
| field | value |
|---|---|
| architecture | diffusers UNet2DModel, 18.5 M params |
| label coding | onehot, 7 channels, 7 codes |
| diffusion | DDPM T=1000, linear beta, v_prediction, zero-terminal-SNR True, offset noise 0.0 |
| sampling | DDIM 50 steps, eta 0.0 |
| resolution | trained 128, exported 512 (INTER_NEAREST) |
| split | block fold 1, train_sha1 eb0aebdd9919 |
| model selection | kl (sampled class-share KL), not the denoising loss |
| River oversampling | 0.0 |
| seed | 42 |
| epochs | 1399 |
| KL(real||gen) | 0.00001 |
| ambiguity rate | 0.00027 |
| recon acc @ t=500 | 0.9789 |
genmask{i}.png, single channel, 512x512, values {0..5} + 255 for ignore.
| class | real % | gen % | ratio |
|---|---|---|---|
| Residential | 52.53 | 52.56 | 1.00 |
| Road | 1.43 | 1.42 | 0.99 |
| River | 0.47 | 0.45 | 0.95 |
| Forest | 12.40 | 12.39 | 1.00 |
| UnusedLand | 14.27 | 14.27 | 1.00 |
| Agricultural | 18.51 | 18.54 | 1.00 |
| Ignore | 0.39 | 0.38 | 0.97 |
| class | frag gen | frag real | ratio | verdict |
|---|---|---|---|---|
| Road | 12.173 | 7.882 | 1.54 | PASS |
| River | 13.884 | 9.123 | 1.52 | PASS |