Research map / Image generation / Fast diffusion sampling

GeoSPRINT: Geometric Redundancy-Aware Step Pruning for Inference in Diffusion Trajectories

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 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 GeoSPRINT, a training-free method that builds non-uniform diffusion sampling schedules by pruning geometrically redundant steps of reference denoising trajectories via a hyperplanarity test, plus a trajectory projection score for straightness.

Problem

Diffusion sampling needs many sequential NFEs; existing training-free accelerators use uniform step-skipping or local error-based step sizing, lacking a global trajectory-level criterion for step importance.

Approach

Generalizes geometric instance reduction (collinearity/coplanarity tests on PCA-projected sequences) to arbitrary d-dimensional latent trajectories via a QR-based hyperplanarity residual test, aggregates retention frequency into a curvature density, and blends it with log-SNR spacing to place steps; also defines a residual-variance trajectory projection score.

Key results

• On CIFAR-10, GeoSPRINT improves FID over DDIM by 0.7-1.1 across 49-89 NFEs and beats DPM-Solver++ at NFE>=30 (e.g., 15.95 vs 17.41 at NFE 49).
• On LSUN Church, FID drops from 1.48 to 1.26 at 52 NFE.
• On Stable Diffusion v1.5 (FID vs 50-step DDIM reference), improvements reach 1.93 FID at NFE 44 (3.45 vs 5.38).

Surprising findings

• First-order DDIM with GeoSPRINT schedule outperforms second-order DPM-Solver++ at NFE>=30 on CIFAR-10.
• Pure log-SNR spacing (beta=0) is worse than DDIM uniform on CIFAR-10 at the tested budgets.

Limitations

• Requires B full reference trajectories offline (e.g., 20,000 NFEs for CIFAR-10), overhead dominates for single-sample generation.
• Blend parameter beta=0.6 tuned on CIFAR-10 only; per-dataset ablations missing for Church and SD v1.5.
• SD v1.5 FID measured against a 50-step DDIM reference rather than ground-truth data; CFG scale fixed at 7.5.

Experimental setup

Datasets: CIFAR-10 LSUN Church COCO 2017 validation (prompts)

Benchmarks: CIFAR-10 LSUN Church Stable Diffusion v1.5 (COCO prompts)

Metrics: FID clean-FID trajectory projection score

Baselines: DDIM DPM-Solver++ DDIM-200

Builds on: DDIM DPM-Solver++ geometric data instance reduction SDM

Closest papers by meaning

• Test-time Scaling of Diffusions with Flow Maps (NeurIPS 2026 · similarity 0.89)
• Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.88)
• Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.88)
• Geometric Trajectory Diffusion Models (NeurIPS 2024 · similarity 0.88)
• Tortoise and Hare Guidance: Accelerating Diffusion Model Inference with Multirate Integration (NeurIPS 2025 · similarity 0.88)
• SCoT: Unifying Consistency Models and Rectified Flows via Straight-Consistent Trajectories (NeurIPS 2025 · similarity 0.88)