mally-2000/saii-cldm-synthetic

SAII-CLDM Synthetic Weights

Quick Start: Click the badge above to run directly in Google Colab (no setup required).

Pretrained synthetic-data model weights for SAII-CLDM in Diffusers format. This repository corresponds to the synthetic-data experiments in the paper. The field-data model was trained separately and is not included here.

Full project code: GeoAI-INV/SAII-CLDM
Paper: arXiv:2506.13529

Load From Hugging Face

Default pipeline: SAII-CLDM.

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "mally-2000/saii-cldm-synthetic",
    custom_pipeline="mally-2000/saii-cldm-synthetic",
    torch_dtype=torch.float32,
    trust_remote_code=True,
).to("cuda")

SAII-LDDPM is also provided as a baseline:

pipe = DiffusionPipeline.from_pretrained(
    "mally-2000/saii-cldm-synthetic",
    custom_pipeline="pipeline_lddpm",
    torch_dtype=torch.float32,
    trust_remote_code=True,
).to("cuda")

The inference algorithms are implemented in codes/pipeline.py. The root pipeline.py, pipeline_cldm.py, and pipeline_lddpm.py files are thin entry points used by Diffusers remote loading.

Run The Bundled Demo

git clone https://huggingface.co/mally-2000/saii-cldm-synthetic
cd saii-cldm-synthetic
pip install -r requirements.txt

python infer.py CLDM  # SAII-CLDM
python infer.py       # SAII-LDDPM baseline

infer.py uses the bundled Overthrust sample and writes outputs under outputs/infer_LDDPM/ or outputs/infer_CLDM/.

Add --eval to run the full bundled Overthrust evaluation:

python infer.py CLDM --eval

Overthrust Results

Impedance-domain metrics on the bundled Overthrust setting. The SAII-CLDM row uses the tuned default sampling parameters in codes/pipeline.py: eta=0, interval=10, sigma_a=15, resampling over the first two-thirds of the DDIM trajectory, and 200 pixel-optimization iterations per resampling step.

Method	Steps	PSNR	SSIM	PCC	RRE
SAII-LDDPM	1000	33.4413	0.9554	0.9957	0.0324
SAII-CLDM	30	33.6184	0.9524	0.9948	0.0321