Research map / Image generation / Fast diffusion sampling

ManifoldCache: Training-Free Diffusion Acceleration via Constraint Manifold Caching

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 ManifoldCache, a training-free, data-free feature-caching schedule for constraint-manifold diffusion models that caches only after the noise-schedule midpoint and assigns deeper blocks longer strides, reporting 1.2-1.9x speedups across eight scientific CMDMs.

Problem

Diffusion models generating samples on hard geometric constraint manifolds (medical volumes, molecules, proteins, crystals, 3D scenes) are slow, needing hundreds to thousands of NFEs, and existing acceleration methods (quantization, pruning, fast solvers, feature caching) either break constraints or

Approach

Derive a closed-form caching schedule from the orthogonal decomposition of the conditional score: caching is enabled only once signal energy exceeds noise energy (alpha-bar >= 1/2), and block strides grow with depth, requiring no calibration, training, or constraint evaluation at inference.

Key results

• On Med-DDPM, ManifoldCache (T*)(2,4) reaches FID 6.93 at 91 s versus 165 s for full inference (1.8x faster), beating DeepCache (20.14) and FasterDiffusion (13.37) on FID.
• On Proteina, ManifoldCache (T*)(2,4) achieves scRMSD 2.02 at 48.18 s vs 1.93 at 90.47 s for full inference, while TeaCache and TaylorSeer reach 8.07 and 9.11.
• On DemoDiff, (T*)(2,3) retains 0.717 design Score vs 0.805 for full inference (about 89%) at 886 s vs 1323 s, while FORA, ToCa, and dKV-Cache drop to 0.44-0.61.

Surprising findings

• Enabling caching across the whole schedule, including the high-noise regime ((0) ablation), roughly triples Med-DDPM FID (6.93 to 21.41), supporting the claim that early caching is harmful.
• ManifoldCache uses more VRAM than full inference on several models (up to 36% more on Med-DDPM), despite being a caching method.

Limitations

• The safe-caching theory is proven only for a two-mode Gaussian mixture, which the evaluated multimodal non-Gaussian models do not necessarily satisfy.
• The depth-monotone result is conditional on an assumed regime condition kappa_t < kappa* that is not derived from the noise schedule.
• Results are reported as single runs without error bars; SV4D 2.0 and Lyra use full inference as reference rather than ground truth.

Experimental setup

Datasets: BraTS 2021 Objaverse GEOM-QM9 PepBench MP-20

Benchmarks: BraTS 2021 3D FID GEOM-QM9 COV-R/COV-P PepBench AAR/C-RMSD/L-RMSD MP-20 Match Rate/RMSD Objaverse PSNR/SSIM/LPIPS DemoDiff 33-task benchmark

Metrics: FID scRMSD COV-R AAR Match Rate PSNR SSIM LPIPS wall-clock time

Baselines: DeepCache FasterDiffusion SmoothCache FORA TeaCache TaylorSeer AB-Cache PFC ToCa dKV-Cache PAB AdaCache

Builds on: DeepCache SmoothCache FORA TeaCache DDIM

Closest papers by meaning

• Rao-Blackwell Gradient Estimators for Equivariant Denoising Diffusion (NeurIPS 2025 · similarity 0.87)
• Manifold Diffusion Fields (ICLR 2024 · similarity 0.87)
• Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models (NeurIPS 2026 · similarity 0.87)
• Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction (ICLR 2025 · similarity 0.86)
• Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.86)
• SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.86)