Path-Guided Flow Matching for Dataset Distillation

What came before

Diffusion-guided methods (Minimax, MGD3, DiT-based sampling) generate through a long stochastic denoising chain and steer it with heuristic guidance terms (mode-clustering, representativeness) applied at every one of many timesteps, which the paper identifies as slow and prone to trajectory instability at low IPC or under strong guidance.

The problem

Heuristic guidance and prototype assignment inside a stochastic diffusion chain make sampling slow (tens of denoising steps, each guided) and unreliable: trajectories can wander off their assigned prototype or collapse toward it too aggressively, hurting downstream generalization, especially when guidance is strong or the per-class sample budget (IPC) is low.

The idea

Replace the stochastic diffusion chain with class-conditional flow matching, which transports noise to data along (near-)straight, ODE-solvable paths in a few deterministic steps, and add a lightweight control term that steers each trajectory continuously toward its assigned class prototype without fighting the model's own learned velocity field, so guidance is both cheaper (fewer steps) and more reliable (a bounded correction to a deterministic path) than diffusion-time guidance.

How it works

A pretrained GMFlow (Gaussian Mixture Flow matching) model (gmflow_imagenet_k8_ema, trained on ImageNet, frozen -- training-free at the generator level) operates in the latent space of a frozen Stable- Diffusion VAE. Per-class prototypes $\{\mu_{y,k}\}_{k=1}^K$ (K=IPC) are found by K-means++ clustering of the class's real-image latent pool. At sampling time, a prototype control velocity is added to the model's own flow field: $u_{proto}(t,z(t),y,k)=\lambda(\mu_{y,k}-\hat z_1(t))$, where $\hat z_1(t)$ is the model's current denoised (endpoint) estimate. A trust-region scale $\alpha(t)$ caps this correction's norm relative to the base velocity's norm so it never dominates, guidance is gated off after $t= t_{end}=0.6$ of the trajectory (early stopping) to avoid distorting fine detail, and initialization is warm-started by interpolating the initial latent toward the assigned prototype by $\eta_{init}$ (tuned per IPC: 0.09/0.05/0.01 for IPC 10/20/50). Only pixels are stored as the distilled artifact.

Evidence

ImageNette 256x256, hard-label (Table 1): IPC10, ConvNet-6: 63.0+/-1.1 vs MGD3 58.5+/-2.1; IPC20, ResNet-18: 70.2+/-0.7 vs MGD3 68.1+/-2.1; IPC50, ResNet-AP: 79.2+/-1.3 vs MGD3 79.1+/-1.5 (roughly tied). ImageNet100 (Table 2), IPC10, ConvNet-6: 25.7+/-0.4 vs MGD3 23.4+/-0.9; IPC20, ResNet-18: 32.9+/-0.5 vs MGD3 32.0+/-1.4. ImageNet-1K IPC10 (Table 3): ConvNet-6 14.1 vs MGD3 13.6; ResNet-18 22.8 vs MGD3 21.5. Also compared against DM, DiT, Minimax, and an unguided flow-matching (FM) baseline; results are reported as newly run under the paper's own protocol rather than quoted verbatim. Efficiency (Table 4): PGFM uses 32 NFE vs MGD3's 50 and Minimax's 50, 2.27 TFLOPs vs MGD3's 17.30 and Minimax's 23.73 -- "10.5x" efficiency gain over Minimax cited in-table, "7.6x" cited in the abstract for the overall diffusion-based comparison. Mode coverage / hit rate (Table 7): FM 10%, MGD3 30%, PGFM (all components) 78%. Component ablation (Table 6, ImageNette IPC10): FM baseline (no guidance) 60.5+/-2.1; + path control 61.6+/-1.2; + early stop 62.1+/-1.3; + trust region 63.0+/-1.1 (best) -- each guidance component adds incrementally. Warm-start sweep (Table 8): eta_init=0 gives 33% hit rate, eta_init=0.09 gives 78%, eta_init=0.15 gives 86% hit rate but lower downstream accuracy (62.0), showing hit rate and accuracy are not monotonically aligned and the paper picks a value short of maximum hit rate.

Limitations

Stated (Appendix E): performance is bounded by the quality and class-coverage of the underlying pretrained GMFlow model and VAE latent space -- a target-trained-generator dependency in a different form (the flow model must already know the classes it generates), since GMFlow was itself pretrained on ImageNet; the method has several hyperparameters ($\lambda$, $t_{end}$, $\rho(t)$, $\eta_{init}$) whose optimal values vary with dataset scale, resolution and class count; prototype discovery uses simple K-means in latent space, noted as potentially suboptimal for fine-grained or long-tailed classes. Observed: not compared against D3HR, CaO2, IGD, CoDA, DAP, ManifoldGD or any training-free or VLM- conditioned diffusion-guided method, only against DM, DiT, Minimax and MGD3, so its standing against the 2025-2026 diffusion frontier (not just the earlier fine-tuned-prior generation) is untested; the ImageNet-1K result is reported at IPC10 only, no IPC50 ImageNet-1K number given.