FlowGuard: Guarding Flow Matching via Conformal Sampling

Ziyun Li, Henrik Boström
Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 266:775-777, 2025.

Abstract

Despite achieving state-of-the-art performance on average, iterative generative models such as diffusion and flow matching remain vulnerable to per-sample failures, sporadically hallucinating object parts, or disregarding conditioning inputs. These rare but critical errors often go undetected, necessitating repeated sampling. We introduce FlowGuard, a lightweight, model-agnostic conformal framework that enforces sample-level reliability during inference. By monitoring trajectory curvature via velocity jumps and rejecting trajectories exceeding a calibrated threshold, FlowGuard provides formal reliability guarantees with negligible computational cost. It operates entirely on cached model outputs and requires no architectural changes. Experiments on CIFAR-10 demonstrate that FlowGuard improves sample quality, reducing FID by up to 2.8%, while enabling early termination of low-quality generations.

Cite this Paper

BibTeX
@InProceedings{pmlr-v266-li25a, title = {FlowGuard: Guarding Flow Matching via Conformal Sampling}, author = {Li, Ziyun and Bostr\"{o}m, Henrik}, booktitle = {Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {775--777}, year = {2025}, editor = {Nguyen, Khuong An and Luo, Zhiyuan and Papadopoulos, Harris and Löfström, Tuwe and Carlsson, Lars and Boström, Henrik}, volume = {266}, series = {Proceedings of Machine Learning Research}, month = {10--12 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v266/main/assets/li25a/li25a.pdf}, url = {https://proceedings.mlr.press/v266/li25a.html}, abstract = {Despite achieving state-of-the-art performance on average, iterative generative models such as diffusion and flow matching remain vulnerable to per-sample failures, sporadically hallucinating object parts, or disregarding conditioning inputs. These rare but critical errors often go undetected, necessitating repeated sampling. We introduce FlowGuard, a lightweight, model-agnostic conformal framework that enforces sample-level reliability during inference. By monitoring trajectory curvature via velocity jumps and rejecting trajectories exceeding a calibrated threshold, FlowGuard provides formal reliability guarantees with negligible computational cost. It operates entirely on cached model outputs and requires no architectural changes. Experiments on CIFAR-10 demonstrate that FlowGuard improves sample quality, reducing FID by up to 2.8%, while enabling early termination of low-quality generations.} }
Endnote
%0 Conference Paper %T FlowGuard: Guarding Flow Matching via Conformal Sampling %A Ziyun Li %A Henrik Boström %B Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2025 %E Khuong An Nguyen %E Zhiyuan Luo %E Harris Papadopoulos %E Tuwe Löfström %E Lars Carlsson %E Henrik Boström %F pmlr-v266-li25a %I PMLR %P 775--777 %U https://proceedings.mlr.press/v266/li25a.html %V 266 %X Despite achieving state-of-the-art performance on average, iterative generative models such as diffusion and flow matching remain vulnerable to per-sample failures, sporadically hallucinating object parts, or disregarding conditioning inputs. These rare but critical errors often go undetected, necessitating repeated sampling. We introduce FlowGuard, a lightweight, model-agnostic conformal framework that enforces sample-level reliability during inference. By monitoring trajectory curvature via velocity jumps and rejecting trajectories exceeding a calibrated threshold, FlowGuard provides formal reliability guarantees with negligible computational cost. It operates entirely on cached model outputs and requires no architectural changes. Experiments on CIFAR-10 demonstrate that FlowGuard improves sample quality, reducing FID by up to 2.8%, while enabling early termination of low-quality generations.
APA
Li, Z. & Boström, H.. (2025). FlowGuard: Guarding Flow Matching via Conformal Sampling. Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 266:775-777 Available from https://proceedings.mlr.press/v266/li25a.html.

Related Material