Research map / Image generation / Fast diffusion sampling

Backbone-Equated Diffusion OOD via Sparse Internal Snapshots

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 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 an MBE protocol for fair cross-backbone diffusion OOD comparison and CFS, a training-free detector that pools a few sparse internal activations of a frozen diffusion backbone at low-noise levels and scores them with an ID-only diagonal Gaussian head.

Problem

Diffusion-based OOD detectors are hard to compare fairly because conclusions depend on backbone family, corruption parameterization, and hidden test-time budget; it is also unclear where in a frozen diffusion backbone the OOD signal resides.

Approach

Align canonical logSNR corruption levels and logical forward-pass cost across diffusion backbones (MBE), then probe a frozen backbone with a few native encoder/decoder block activations at one low-noise level (CFS), scored by an ID-only diagonal statistic, rather than output-space summaries.

Key results

• On CIFAR-scale MBE benchmark, CFS(1x2) achieves AvgAUROC 0.887 (improved) and 0.916 (EDM) with 1 forward pass, versus 0.792/0.796 for MSMA at 10 forwards.
• CFS(1x2) improves AvgWorstAUROC from 0.689 to 0.814 on EDM and from 0.688 to 0.799 on improved-diffusion relative to the strongest harmonized output-space baselines.
• On the external positioning table, CFS(1x2)-ImageNet reports Avg AUROC 0.962 at 1 forward, above DiffPath-6D-CelebA (0.931) at 10 forwards.

Surprising findings

• A single late-decoder snapshot (CFS_dec(1x1)) is nearly as strong as the paired encoder-decoder variant.
• Under the EDM backbone, CFS_dec(1x1) reaches AvgAUROC 0.919, slightly above CFS(1x2) at 0.916.

Limitations

• Requires internal access to the diffusion backbone, so less black-box than output-space scores.
• Internal features remain architecture- and source-dependent; portability is structural rather than representational.
• Theory is local and diagnostic rather than universal; hard multimodal ID regimes such as CIFAR-10 remain challenging.

Experimental setup

Datasets: CIFAR-10 SVHN CelebA32 CIFAR-100 DTD ImageNet-64 NINCO SSB-hard Textures

Benchmarks: CIFAR-scale OOD benchmark (12 ID-to-OOD pairs per backbone) ImageNet200 / ImageNet1K (appendix)

Metrics: AUROC AvgAUROC AvgWorstAUROC FPR95 forward passes per image Spearman correlation

Baselines: MSMA DiffPath DDPM-OOD GEPC SCOPED DLSR

Builds on: DDPM (improved-diffusion) EDM OpenOOD

Closest papers by meaning

• GOOD: Training-Free Guided Diffusion Sampling for Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.87)
• Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.86)
• SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.86)
• One More Step: A Versatile Plug-and-Play Module for Rectifying Diffusion Schedule Flaws and Enhancing Low-Frequency Controls (CVPR 2024 · similarity 0.86)
• SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.85)
• EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.85)