ReflexFlow: Rethinking Learning Objective for Exposure Bias Alleviation in Flow Matching
On the map: Diffusion models → Flow matching, time series and science
Abstract (arXiv)
Despite tremendous recent progress, Flow Matching methods still suffer from exposure bias due to discrepancies in training and inference. This paper investigates the root causes of exposure bias in Flow Matching, including: (1) the model lacks generalization to biased inputs during training, and (2) insufficient low-frequency content captured during early denoising, leading to accumulated bias. Based on these insights, we propose ReflexFlow, a simple and effective reflexive refinement of the Flow Matching learning objective that dynamically corrects exposure bias. ReflexFlow consists of two components: (1) Anti-Drift Rectification (ADR), which reflexively adjusts prediction targets for biased inputs utilizing a redesigned loss under training-time scheduled sampling; and (2) Frequency Compensation (FC), which reflects on missing low-frequency components and compensates them by reweighting the loss using exposure bias. ReflexFlow is model-agnostic, compatible with all Flow Matching frameworks, and improves generation quality across datasets. Experiments on CIFAR-10, CelebA-64, and ImageNet-256 show that ReflexFlow outperforms prior approaches in mitigating exposure bias, achieving a 35.65% reduction in FID on CelebA-64.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Flow matching, time series and science, one of 14 directions in Diffusion models. Builds on 44 of its 58 references that are in the map. First posted on arXiv 4 Dec 2025.
Builds on
44 of 58 references.
Closest ancestors · 6
Same direction, newest first
- 2025CurveFlow: Curvature-Guided Flow Matching for Image Generation20 Aug
- 2025SplitMeanFlow: Interval Splitting Consistency in Few-Step Generative Modeling22 Jul
- 2025Improving Progressive Generation with Decomposable Flow Matching24 Jun
- 2025Align Your Flow: Scaling Continuous-Time Flow Map Distillation17 Jun
- 2022Flow Matching for Generative Modeling6 Oct
- 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow7 Sep
Other references · 38
- 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
- 2020Denoising Diffusion Probabilistic Modelscited by 21,076
- 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,305
- 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,863
- 2015Deep Unsupervised Learning using Nonequilibrium Thermodynamicscited by 7,058
- 2022Scalable Diffusion Models with Transformerscited by 6,054
- 2019Generative Modeling by Estimating Gradients of the Data Distributioncited by 4,101
- 2014Deep Learning Face Attributes in the Wildcited by 3,693
- 2016Improved Techniques for Training GANscited by 2,970
- 2025Wan: Open and Advanced Large-Scale Video Generative Modelscited by 2,708
- 2023Consistency Modelscited by 1,822
- 2022Building Normalizing Flows with Stochastic Interpolantscited by 995
26 more
- 2015Scheduled Sampling for Sequence Prediction with Recurrent Neural Networkscited by 803
- 2024SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformerscited by 737
- 2015Sequence Level Training with Recurrent Neural Networkscited by 576
- 2025Mean Flows for One-step Generative Modelingcited by 561
- 2023Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusioncited by 459
- 2024One Step Diffusion via Shortcut Modelscited by 391
- 2025Kimi-Audio Technical Reportcited by 222
- 2024FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Modelscited by 217
- 2019Bridging the Gap between Training and Inference for Neural Machine Translationcited by 123
- 2019Generalization in Generation: A closer look at Exposure Biascited by 123
- 2023Elucidating the Exposure Bias in Diffusion Modelscited by 101
- 2023FreeInit: Bridging Initialization Gap in Video Diffusion Modelscited by 95
- 2023Input Perturbation Reduces Exposure Bias in Diffusion Modelscited by 90
- 2023Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Stepscited by 73
- 2023Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistentcited by 58
- 2023PLANNER: Generating Diversified Paragraph via Latent Language Diffusion Modelcited by 58
- 2024Video-Guided Foley Sound Generation with Multimodal Controlscited by 49
- 2023Exploiting the Signal-Leak Bias in Diffusion Modelscited by 36
- 2023On Error Propagation of Diffusion Modelscited by 31
- 2024Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Modelcited by 28
- 2024MotiF: Making Text Count in Image Animation with Motion Focal Losscited by 10
- 2025Model Already Knows the Best Noise: Bayesian Active Noise Selection via Attention in Video Diffusion Modelcited by 10
- 2022Markup-to-Image Diffusion Models with Scheduled Samplingcited by 6
- 2025Proteus-ID: ID-Consistent and Motion-Coherent Video Customizationcited by 6
- 2025Frequency Regulation for Exposure Bias Mitigation in Diffusion Modelscited by 6
- 2025FastInit: Fast Noise Initialization for Temporally Consistent Video Generationcited by 3
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Exposure Bias Can Alleviate Itself via Directional and Frequency Rectification in Flow Matching
- 2025Contrastive Flow Matching
- 2024Flow Matching Posterior Sampling: A Training-free Conditional Generation for Flow Matching
- 2025MixFlow Training: Alleviating Exposure Bias with Slowed Interpolation Mixture
- 2026Posterior Augmented Flow Matching
- 2026FastFlow: Accelerating The Generative Flow Matching Models with Bandit Inference
- 2026Noise-Robust Conditional Flow Matching: Generating Clean Samples from Noisy Datasets
- 2026Training-Free Refinement of Flow Matching with Divergence-based Sampling