Research map / Image generation / Fast diffusion sampling

Path-Guided Flow Matching for Dataset Distillation

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Flow matching distillation · 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 PGFM, a flow-matching dataset distillation method that samples class-conditional latents from a pretrained GMFlow generator with early-stopped, trust-region prototype guidance and warm start, reporting gains over diffusion-based distillation baselines at lower sampling cost.

Problem

Diffusion-based generative dataset distillation relies on long denoising chains and heuristic guidance or prototype assignment, which raises sampling cost and makes prototype-consistent control unstable, especially at low IPC.

Approach

Use a pretrained flow-matching generator (GMFlow) in a frozen VAE latent space with deterministic ODE sampling, plus lightweight k-means++ prototype guidance applied only early in the trajectory, bounded by a trust region, and a warm start toward assigned prototypes.

Key results

• On ImageNette IPC=10 with ConvNet-6, PGFM reaches 63.0% vs 58.5% for MGD$^3$ (reported in table).
• Prototype hit rate rises from 30% (MGD$^3$) and 10% (FM) to 78% for PGFM on ImageNette IPC=10.
• On ImageNet-1K IPC=10, PGFM improves over MGD3 by +0.5, +2.7, and +1.3 points on ConvNet-6, ResNet18, and ResNet-AP.

Surprising findings

• Unguided flow-matching sampling alone already outperforms diffusion-based distillation baselines such as MGD$^3$ and MinimaxDiffusion on ImageNet100 at IPC=10.
• Over-strong prototype guidance washes out details and produces blurred or gray images.

Limitations

• Performance bounded by the quality and class coverage of the pretrained GMFlow and VAE.
• Several hyperparameters (lambda, s_end, rho, eta_init) need dataset-specific tuning.
• Prototype discovery uses simple k-means, which may be suboptimal for fine-grained or long-tailed classes.

Experimental setup

Datasets: ImageNette ImageIDC ImageNet100 ImageNet-1K

Benchmarks: ImageNette ImageIDC ImageNet100 ImageNet-1K

Metrics: top-1 accuracy hit rate FLOPs parameters NFE

Baselines: DM DiT MinimaxDiffusion MGD$^3$ FM (GMFlow unguided) Random Herding IDC-1 MTT DSDM

Builds on: GMFlow MGD$^3$ MinimaxDiffusion Distribution Matching

Closest papers by meaning

• Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.91)
• Efficient Dataset Distillation for Pre-Trained Self-Supervised Models via Statistical Flow Matching (NeurIPS 2026 · similarity 0.90)
• Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.90)
• How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · similarity 0.90)
• Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
• Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.90)