Research map / Image generation / Fast diffusion sampling

BGDB: Bernoulli-Gaussian Decision Block with Improved Denoising Diffusion Probabilistic Models

AAAI 2025 · Image generation › Fast diffusion sampling · approach: Distributional control of diffusion · 1 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 the Bernoulli-Gaussian Decision Block (BGDB), an auxiliary IDDPM-based training-only module that models logits as Gaussian approximations of multiple Bernoulli trials to supervise a single training run, reporting small gains on segmentation and classification.

Problem

Discriminative classifiers' logits from a single training pass may be unstable and imprecise estimates of true class probability; generative models are typically used only for feature augmentation without a solid theoretical basis.

Approach

Treat each forward pass as a Bernoulli trial and use an Improved DDPM attached to the logits during training to produce a Gaussian (mean and variance) estimate approximating the mean of many trials (via CLT / De Moivre-Laplace), supervising the backbone with BCE/MSE and hybrid diffusion losses; the block is removed at inference.

Key results

• On Cityscapes, DeepLabV3+ with BGDB improved mIoU by 0.08 to 1.48 points; on Pascal VOC by 0.21 to 0.41 points.
• On ISIC skin lesion segmentation, BGDB raised Dice for UNETR, FCN, U-Net, ResUNet and Attention U-Net by 0.62 to 4.74 points, with U-Net++ dropping 0.29.
• On ISIC skin lesion classification, BGDB improved accuracy by 0.54 to 1.60 points for most backbones (ViT dropped 0.04) and AUC by 0.24 to 0.74 points.

Measured effects reported in the paper

Surprising findings

• The U-Net++ backbone showed a slight decline (-0.29 Dice) with the block, and ViT accuracy dropped slightly (-0.04).

Limitations

• Theoretical justification relies on a sufficiently large number of trials n that is not established in practice.
• IDDPM training is slow and parameter-heavy, limiting diffusion time steps (minimum 25) and making 3D use impractical.
• Classification with the U-Net encoder-decoder needs awkward logit resizing and averaging, and gains are modest.

Experimental setup

Datasets: Cityscapes Pascal VOC ISIC 2018

Benchmarks: Cityscapes Pascal VOC ISIC 2018

Metrics: mIoU Dice score accuracy AUC

Baselines: DeepLabV3+ (MobileNet) DeepLabV3+ (ResNet101) DeepLabV3+ (HRNet) U-Net Attention U-Net U-Net++ FCN ResUNet UNETR DenseNet169 ViT ResNet50

Builds on: DDPM Improved DDPM DeepLabV3+ Dice loss

Closest papers by meaning

• Sampling Binary Data by Denoising through Score Functions (ICML 2025 · similarity 0.89)
• Bernoulli Flow Models: Self-Consistent Generative Modeling for Binary Data (NeurIPS 2026 · similarity 0.87)
• Classification Diffusion Models: Revitalizing Density Ratio Estimation (NeurIPS 2024 · similarity 0.87)
• Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (ICML 2025 · similarity 0.87)
• Glauber Generative Model: Discrete Diffusion Models via Binary Classification (ICLR 2025 · similarity 0.87)
• Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds (NeurIPS 2025 · similarity 0.86)