Research map / Image generation / Fast diffusion sampling

Tempered Guided Diffusion

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 Tempered Guided Diffusion, an annealed SMC sampler over clean-signal tempered posteriors for training-free conditional diffusion, plus an accelerated pruning variant (A-TGD) that improves matched-runtime image inverse-problem reconstruction.

Problem

Training-free conditional diffusion samplers allocate computation inefficiently across independent guided trajectories, which vary widely in quality and cannot recover from poor early decisions.

Approach

Treat reconstructions from noisy diffusion states as an interacting particle population targeting annealed clean-space posteriors p(x0)p(y|x0)^λ, reweighting by incremental likelihood ratios and resampling; A-TGD prunes to one high-likelihood trajectory partway through sampling.

Key results

• On FFHQ inpainting, A-TGD reaches 25.14 PSNR / 0.853 SSIM / 0.122 LPIPS vs DAPS (N=4) 24.52 / 0.790 / 0.157 at matched wall-clock.
• On FFHQ phase retrieval, A-TGD achieves 29.43 PSNR and 0.172 LPIPS vs DAPS (N=4) 27.81 and 0.207.
• On ImageNet phase retrieval, A-TGD matches DAPS (N=4) in PSNR (22.18 vs 22.41 for DAPS N=4 is lower, so not matched) and does not improve LPIPS (0.326 vs 0.325).

Surprising findings

• On ImageNet phase retrieval, A-TGD does not beat DAPS (N=4) on LPIPS and PSNR, showing gains are dataset-dependent.
• In the 2D experiment, DPS and DPS-DAPS do not benefit from additional independent particles, while TGD SWD decreases with particle count.

Limitations

• Consistency theorem applies only to idealized exact-reconstruction, unpruned TGD; A-TGD and practical solvers lack the guarantee.
• Aggressive early pruning can discard good trajectories in challenging phase retrieval, increasing variance.
• Hybrid reconstruction modules with intermediate tempering did not give consistent improvements in preliminary experiments.

Experimental setup

Datasets: FFHQ ImageNet

Benchmarks: FFHQ inpainting FFHQ phase retrieval ImageNet inpainting ImageNet phase retrieval controlled 2D inverse problem

Metrics: PSNR SSIM LPIPS sliced Wasserstein distance

Baselines: DPS DAPS (N=1) DAPS (N=4) DPS-DAPS ablation

Builds on: DAPS MPGD DPS TDS EDM

Closest papers by meaning

• Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.90)
• Multilevel and Sequential Monte Carlo for Training-Free Diffusion Guidance (NeurIPS 2026 · similarity 0.90)
• DEFT: Efficient Fine-tuning of Diffusion Models by Learning the Generalised $h$-transform (NeurIPS 2024 · similarity 0.90)
• Provable Posterior Sampling with Denoising Oracles via Tilted Transport (NeurIPS 2024 · similarity 0.89)
• Guided Diffusion Sampling on Function Spaces with Applications to PDEs (NeurIPS 2025 · similarity 0.89)
• A Mixture-Based Framework for Guiding Diffusion Models (ICML 2025 · similarity 0.89)