Research map / Image generation / Autoregressive and latent generation

A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset

AAAI 2025 · Image generation › Autoregressive and latent generation · approach: Synthetic training data from generators · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UB-Diff, a diffusion model with a one-in-two-out encoder-decoder that builds a co-latent space so paired seismic waveform and velocity map data can be generated from unbalanced training data.

Problem

Generating paired multi-modal scientific data (seismic waveforms and velocity maps) when one modality is abundant and the paired counterpart is scarce, a setting existing multi-modal diffusion methods that require fully paired data do not handle.

Approach

A one-in-two-out encoder-decoder maps the abundant modality (velocity maps) into a shared co-latent space with two decoders (transformer for seismic, CNN for velocity); a two-step training scheme first trains on all velocity maps self-supervised, then fine-tunes on the few paired samples, and a latent diffusion model is trained on the abundant modality's latents to generate pairs.

Key results

• On FlatVel-A with 5k paired seismic data, UB-Diff reaches FID 15.64 for velocity maps and 9.25 for seismic vs. MT-Diffusion's 99.89 and 71.53.
• With 1k paired seismic data on FlatVel-A, UB-Diff gives FID 16.45 (velocity) and 97.03 (seismic), versus MT-Diffusion's 336.54 and 318.29.
• Training InversionNet on 10k UB-Diff generated pairs improves SSIM on FlatVel-A from 0.6511 (MT-Diffusion) to 0.7881 at 1k seismic samples.

Measured effects reported in the paper

Surprising findings

• MT-Diffusion performs poorly on Style-A, which the authors attribute to the decomposition loss interfering with the diffusion loss.
• UB-Diff w/o the two-step optimization sometimes gives better seismic FID on Style-A than the full UB-Diff, and the freezing choice involves a trade-off between the two decoders.

Limitations

• Evaluated only on OpenFWI synthetic geophysics datasets; no real-world field seismic data.
• UB-Diff does not consistently beat baselines on every dataset or metric (e.g., seismic FID on Style-A and some SSIM/MAE entries on FlatFault-A and CurveFault-A).
• Architecture and training are more complex and need additional compute, as the authors acknowledge.

Experimental setup

Datasets: OpenFWI FlatVel-A OpenFWI CurveVel-A OpenFWI FlatFault-A OpenFWI CurveFault-A OpenFWI Style-A

Benchmarks: OpenFWI FlatVel-A OpenFWI CurveVel-A OpenFWI FlatFault-A OpenFWI CurveFault-A OpenFWI Style-A

Metrics: FID MAE MSE SSIM

Baselines: MT-Diffusion DDPM Stable Diffusion InversionNet

Builds on: DDPM Latent Diffusion Models MT-Diffusion InversionNet

Closest papers by meaning

• A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations (ICLR 2025 · similarity 0.85)
• Broadband Ground Motion Synthesis by Diffusion Model with Minimal Condition (ICML 2025 · similarity 0.85)
• Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series (NeurIPS 2024 · similarity 0.84)
• Unified Framework for Diffusion Generative Models in SO(3): Applications in Computer Vision and Astrophysics (AAAI 2024 · similarity 0.84)
• DiffATS: Diffusion in Aligned Tensor Space (NeurIPS 2026 · similarity 0.84)
• Isometric Representation Learning for Disentangled Latent Space of Diffusion Models (ICML 2024 · similarity 0.84)