Research map / Image generation / Fast diffusion sampling

Test-time Scaling of Diffusions with Flow Maps

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Fast diffusion ODE solvers · 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 Flow Map Trajectory Tilting (FMTT), which uses a flow map's look-ahead to tilt diffusion/flow sampling toward a reward, with exact importance weights via a Jarzynski/SMC scheme or top-n search, and applies it to FLUX-distilled flow maps with VLM rewards.

Problem

Test-time reward guidance of diffusion/flow models requires a reward defined only on clean terminal samples, so gradient-based tilting needs a look-ahead prediction of the endpoint; denoiser-based look-ahead is inaccurate at early noise levels.

Approach

Use a two-time flow map X_{t,1} as the exact look-ahead inside the reward, tilt the SDE with the gradient of r(X_{t,1}), and show the Jarzynski importance weights reduce to the simple integral of the reward along the look-ahead trajectory.

Key results

• On GenEval, FMTT reaches mean 0.79 versus 0.65 for FLUX.1 [dev] and 0.75 for Best-of-N on FLUX.1 [dev], at 1400 NFE.
• On GenEval, the 1-step denoiser look-ahead variant of FMTT scores 0.75, below the full flow map look-ahead's 0.79.
• On UniGenBench++ with a VLM reward, FMTT with flow-map look-ahead gives larger gains than best-of-N at lower compute, while denoiser look-ahead barely beats best-of-N.

Surprising findings

• Denoiser look-ahead tilting provides little to no improvement over best-of-N on UniGenBench++, attributed to blurry predictions at high noise.

Limitations

• Reward hacking risk when maximizing VLM rewards, requiring carefully written questions
• Gains limited on human-preference rewards because base FLUX is already post-trained on preference data
• Requires a trained flow map model, and the Laplacian/extra computation for some weight variants

Experimental setup

Datasets: MNIST GenEval UniGenBench++

Benchmarks: GenEval UniGenBench++ MNIST tilted sampling

Metrics: GenEval mean score UniGenBench++ overall score NFE average reward thermodynamic length

Baselines: Best-of-N Multi-Best-of-N Beam search ReNO FLUX.1-dev 1-step denoiser look-ahead

Builds on: flow matching stochastic interpolants Align Your Flow ReNO Jarzynski equality

Closest papers by meaning

• Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.90)
• Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.90)
• Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
• Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing (NeurIPS 2025 · similarity 0.89)
• Wasserstein Gradient Flows and Forward-Only Diffusion Are Not Enough for Multimodal Sampling (NeurIPS 2026 · similarity 0.89)
• Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.89)