PositivePassion/openfwi-diffusion-priors

OpenFWI Diffusion Priors

Pretrained unconditional DDPM priors for six OpenFWI velocity-model families. Each model contains a single-channel UNet2DModel and a DDPMScheduler in a Diffusers-compatible directory layout.

Included models

Directory	OpenFWI family
FlatFault-A	FlatFault A
FlatFault-B	FlatFault B
CurveFault-A	CurveFault A
CurveFault-B	CurveFault B
CurveVel-A	CurveVel A
CurveVel-B	CurveVel B

The UNets operate on normalized, single-channel 72 x 72 inputs. The inversion workflow crops the generated result to the 70 x 70 OpenFWI model domain.

Download

python3 -m pip install -U huggingface_hub
hf download PositivePassion/openfwi-diffusion-priors --local-dir models

To download only one family:

hf download PositivePassion/openfwi-diffusion-priors \
  --include "CurveFault-A/*" \
  --local-dir models

Loading

from diffusers import DDPMScheduler, UNet2DModel

model_dir = "models/CurveFault-A"
unet = UNet2DModel.from_pretrained(model_dir, subfolder="unet")
scheduler = DDPMScheduler.from_pretrained(model_dir, subfolder="scheduler")

See SHA256SUMS for weight-file checksums.