Research map / Image generation / Fast diffusion sampling

Observation-Guided Diffusion Probabilistic Models

CVPR 2024 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 3 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 observation-guided diffusion probabilistic models (OGDM), which add a time-conditioned discriminator-based emission loss during training to make diffusion denoisers more accurate for few-step sampling, with no inference overhead.

Problem

Diffusion models sampled with large step sizes (few NFEs) deviate from the Gaussian reverse-process assumption, so the training objective mismatches the inference negative log-likelihood and quality degrades.

Approach

Introduce an observation process at each diffusion step modeled as a Bernoulli emission (whether a noisy state lies on the real-data manifold), implemented by a time-dependent discriminator that judges one-step-projected denoiser predictions against real data; the resulting extra adversarial-style loss is added to the standard denoising loss, and the discriminator is discarded at inference.

Key results

• On CIFAR-10 with ADM at 10 NFEs, FID improves from 15.20 (baseline) to 11.63 with OGDM trained from scratch.
• On CelebA with ADM at 10 NFEs, FID improves from 11.92 to 7.91 with OGDM.
• With Heun's method and EDM on CIFAR-10 at 11 NFEs, FID drops from 14.71 to 13.54 with OGDM fine-tuning; OGDM with S-PNDM also improves few-step FID across datasets.

Measured effects reported in the paper

Surprising findings

• OGDM fine-tuning on EDM at 10 NFEs on CIFAR-10 cuts FID from 19.32 to 9.28, a larger gain than from-scratch training in some settings.
• Gains are not guaranteed when the projection function and sampler are misaligned, e.g. F-PNDM results are mixed.

Limitations

• Training cost is about 80% higher due to the discriminator at equal iterations.
• Benefits depend on alignment between the projection function and the sampler.
• Hyperparameters (k, gamma) were chosen per setup and the observation design is only one of several possible formulations.

Experimental setup

Datasets: CIFAR-10 CelebA LSUN Church

Benchmarks: CIFAR-10 CelebA LSUN Church

Metrics: FID recall

Baselines: ADM EDM S-PNDM F-PNDM DDIM Analytic-DPM FastDPM GENIE CT

Builds on: DDPM DDIM EDM Diffusion-GAN PNDM

Closest papers by meaning

• EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)
• Efficient Denoising Diffusion via Probabilistic Masking (ICML 2024 · similarity 0.89)
• Boosting Diffusion Models with Moving Average Sampling in Frequency Domain (CVPR 2024 · similarity 0.88)
• Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.88)
• A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.88)
• Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models (ICLR 2025 · similarity 0.88)