Research map / Image generation / Fast diffusion sampling

Drift Flow Matching

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 Drift Flow Matching (DFM), which trains a two-time mean-velocity transport map with a drift-based loss on marginal pairs, letting one model generate in one step (like Drift Models) or multiple steps with test-time scaling (like Flow Matching).

Problem

Drift Models give efficient one-step generation but lack a mechanism to improve quality with more inference steps, while Flow Matching and Diffusion scale with NFE but need many steps.

Approach

Learn a two-time transport map T_{t,r}(x_t)=x_t+(r-t)u(x_t,t,r) that moves samples from time t to r along the Flow Matching path, with the drift loss applied to the marginal pair (p_t, p_r) grouped per time pair, so that one step recovers a Drift Model and many steps recover Flow Matching.

Key results

• On ImageNet 256x256 with DFM L/2 (463M), one-step FID 1.52 vs Drifting Model L/2 1.54; FID improves to 1.45 at NFE 2 and 1.31 at NFE 10.
• On FFHQ latent generation, DFM with 5 steps reaches FID 75.9 and EMD 196.2, versus 116.2 FID for one-step Drift Model.
• On robotic control, DFM policy at NFE 10 reaches 0.86 on ToolHang state (vs 0.38 Drift Policy at NFE 1), and 0.65 on BlockPush Phase 1.

Surprising findings

• Using the endpoint-prediction parameterization instead of mean velocity degrades FID at all NFEs in ablation.
• Setting the number of positive/negative samples n_g=1 makes one-step generation break down (FID 220 at NFE 1), since the drift reduces to a conditional velocity target.

Limitations

• Gains at higher NFE on ImageNet are reported only against one-step baselines in some rows; multi-step comparisons are limited.
• Relies on the kernel-based drift field whose identifiability is only heuristic per the paper's own remark.

Experimental setup

Datasets: ImageNet-1k MNIST FFHQ 2D synthetic datasets (F, M, two-moons, checkerboard)

Benchmarks: ImageNet 256x256 MNIST FFHQ robotic control tasks (Lift, Can, ToolHang, PushT, BlockPush, Kitchen)

Metrics: FID IS EMD classification accuracy success rate

Baselines: Flow Matching Mean Flow Drift Model iCT Shortcut MeanFlow AdvFlow iMeanFlow S-VFM iMM DiT-XL/2 SiT-XL/2

Builds on: Drifting Models Flow Matching Mean Flow Diffusion Policy

Closest papers by meaning

• D-Flow: Differentiating through Flows for Controlled Generation (ICML 2024 · similarity 0.89)
• Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
• Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
• Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.88)
• How to build a consistency model: Learning flow maps via self-distillation (NeurIPS 2025 · similarity 0.88)
• Path-Guided Flow Matching for Dataset Distillation (NeurIPS 2026 · similarity 0.88)