Research map / Image generation / Fast diffusion sampling

Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space

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 Temporal Alignment Guidance (TAG), which uses a learned time predictor's gradient as an extra guidance term to pull diffusion samples back onto the timestep-appropriate data manifold, improving fidelity under external guidance, multi-condition guidance, few-step sampling, and text-to-image reward alignment.

Problem

External guidance (training-free guidance, multi-conditional guidance, few-step sampling) pushes diffusion samples off the data manifold into low-density regions where the learned score is unreliable, degrading fidelity.

Approach

Treat timestep as a conditioning variable and add the gradient of a log-probability time predictor, grad_x log p(t|x), as an on-manifold anchor term in each reverse step, with a single-condition or unconditional time predictor reparameterization for multi-condition guidance.

Key results

• On the TFG benchmark, TFG+TAG reduces CIFAR10 FID from 114.1 to 102.7 and raises accuracy from 55.8 to 61.5.
• DAS+TAG raises single-objective Aesthetic reward from 7.948 to 9.087 and lowers Time-Gap from 90.04 to 28.84 on SD v1.5.
• In single-step CIFAR10 few-step generation, TAG lowers FID by 41.1%.

Surprising findings

• Corrector sampling can become ineffective or even degrade sample quality under external guidance.

Limitations

• Once sample fidelity is high, further narrowing the time-gap yields marginal or no gains.
• Time predictor architecture is simple and may limit gains.
• Theoretical off-manifold analysis is informal; full theorem deferred.

Experimental setup

Datasets: CIFAR10 ImageNet CelebA Cat QM9 HPSv2 prompts WikiArt PartiPrompts

Benchmarks: TFG benchmark CIFAR10 ImageNet CelebA QM9 molecular properties audio declipping audio inpainting PartiPrompts style transfer

Metrics: FID LPIPS accuracy FAD DTW MAE stability KID Aesthetic score CLIP score

Baselines: DPS TFG TCS Timestep Guidance Self-Guidance DAS

Builds on: training-free guidance DPS DAS classifier guidance DDIM

Closest papers by meaning

• Local Manifold Approximation and Projection for Manifold-Aware Diffusion Planning (ICML 2025 · similarity 0.88)
• Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models (NeurIPS 2025 · similarity 0.88)
• Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.87)
• What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.87)
• Test-time Alignment of Diffusion Models without Reward Over-optimization (ICLR 2025 · similarity 0.87)
• Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion (ICLR 2025 · similarity 0.87)