Research map / Image generation / Fast diffusion sampling

Towards more general control of diffusion models using Jeffrey Guidance

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance for conditional diffusion · 0 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 Jeffrey guidance, which uses Jeffrey's rule of conditioning to reweight a pretrained diffusion model so that a chosen marginal (embedding or attribute distribution) matches a target, and applies it to Inception-embedding FID reduction and CelebA-HQ attribute fairness.

Problem

Diffusion guidance usually targets class-conditional or heuristically defined distributions; there is no principled way to update a model so that a general marginal (e.g., an embedding distribution or attribute balance/independence) matches an explicit target while minimally perturbing the joint.

Approach

Apply Jeffrey's rule of conditioning to diffusion: reweight p(x) by the density ratio rho(y)=p*(y)/p(y) of a feature y=u(x), and implement it as a guidance term added to the score using a Tweedie-based estimate of x0, with rho estimated by a binary classifier.

Key results

• On CIFAR-10, FID drops from about 5.88 to 2.55 against the training set (50k) and from 9.76 to 6.71 on the test set (10k).
• On FFHQ test set, FID drops from about 20.13 to 12.91 with delta=10 and lambda=30.
• On CelebA-HQ gender parity, lambda=2.0 moves male proportion from 0.296 to 0.522, reducing FD from 0.288 to 0.032; Male-Young correlation reaches near zero (0.005) at lambda=3.0 with FID rising from 22.30 to 23.59.

Surprising findings

• FID can drop substantially with little visible change in generated images, highlighting FID's limits as a perceptual metric.
• Applying guidance only at the last denoising step (delta=10) gave the best FID, while larger guidance windows degraded it.

Limitations

• Guidance uses an approximate x0-hat (Tweedie) estimate, so the sampled distribution is not exact; the exact scheme is only discussed theoretically.
• Density ratio estimation relies on a classifier and requires differentiable feature maps such as Inception.
• Gains in FID do not necessarily reflect perceptual improvement; the lambda and delta values need tuning.

Experimental setup

Datasets: CIFAR-10 FFHQ CelebA-HQ CelebA

Benchmarks: CIFAR-10 FFHQ CelebA-HQ

Metrics: FID Fairness Discrepancy Pearson correlation CMMD

Baselines: standard guidance (ancestral sampling plus class-conditional guidance) unguided model

Builds on: classifier guidance universal guidance DDIM Tweedie's formula ADM

Closest papers by meaning

• Gradient Guidance for Diffusion Models: An Optimization Perspective (NeurIPS 2024 · similarity 0.89)
• Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.88)
• Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.88)
• Variational Control for Guidance in Diffusion Models (ICML 2025 · similarity 0.88)
• Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.87)
• Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.87)