Research map / Image generation / Fast diffusion sampling

Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Guidance and alignment for flow matching · 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 CoFlow, an online neural contextual bandit that adaptively selects the number of integration steps per prompt for flow-matching models, using an unsupervised velocity-variation reward, reporting ~2.5x fewer NFEs on image, editing and video generation.

Problem

Flow matching inference needs many sequential ODE evaluations, and fixed step budgets ignore that different prompts need different amounts of computation; existing acceleration methods either require retraining or use heuristic step skipping.

Approach

Treat step-budget selection as an online contextual bandit: a small MLP (with NeuralUCB exploration) maps prompt features to a step count, rewarded by an intrinsic L1 relative velocity-change surrogate minus a per-step cost, so no external quality model or retraining is needed.

Key results

• On BAGEL, CoFlow reaches GenEval 0.77 at 17.7 NFE versus 0.78 for full 50-step inference.
• On FLUX, CoFlow reaches GenEval 0.65 at 19.5 NFE (Table 1) versus 0.66 at 50 steps; the paper reports speedups over 2.5x.
• The surrogate velocity-variation statistic has mean within-prompt Spearman 0.924 with deviation from a 100-step reference, but only 0.255 cross-prompt at fixed budget.

Surprising findings

• The velocity-variation surrogate tracks within-trajectory discretization sensitivity but is weakly associated with prompt difficulty across prompts at a fixed budget (Spearman 0.255).
• Direct step reduction with a solver-consistent schedule retained better quality than trajectory truncation at comparable compute.

Limitations

• Requires an initial exploration phase, though reported results include exploration samples.
• Effectiveness depends on the expressiveness of the handcrafted and embedding prompt features.
• Surrogate reward is not a direct measure of quality; the theoretical bound relies on regularity assumptions and is not tied to the selected policy.

Experimental setup

Datasets: GenEval GEdit VBench

Benchmarks: GenEval GEdit VBench CLIP-IQA

Metrics: GenEval overall score CLIP-IQA NFE inference time G_SC G_PQ G_O VBench score BRISQUE

Baselines: Full generation TeaCache InstaFlow PeRFlow FlowCast FastFlow AdaDiff

Builds on: NeuralUCB AdaDiff flow matching TeaCache FastFlow

Closest papers by meaning

• Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.90)
• Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.89)
• Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
• Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.88)
• On the Guidance of Flow Matching (ICML 2025 · similarity 0.88)
• Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.88)