Prediction horizons affect image quality depending on timing and state
Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow
Machine LearningArtificial Intelligence
Summary
Prediction horizon refers to how far ahead a model looks when making predictions, usually linked to time intervals. This paper shows that changing the prediction horizon separately from how often the model updates its state can impact the quality of generated images. The authors found that the effect depends not just on timing but also on the states the model receives at each step, and that adjusting prediction horizons can improve or worsen results depending on the setup. They also demonstrated rules that improve image quality without retraining the models by changing how queries to the model are scheduled during inference.
What this means in practice
- •For image generation developers: Improve image synthesis quality by adjusting prediction horizon rules during model inference without retraining.
- •For machine learning engineers: Optimize sampling strategies in diffusion or flow-based models by decoupling update steps from prediction queries to enhance output fidelity.
Authors
Tinglan Wang, Haixu Liu, Shaobo Chen, Nuo Qian, Qi Gao, Kai Wen, Yu Yao
Abstract
MeanFlow predicts interval-average velocities, but conventional few-step sampling ties the interval queried from the model to the duration of the state update. We separate these two quantities at inference, fixing the model, update grid, and number of queries while varying the prediction horizon. Our central finding is that the horizon response depends on both query position and the states entering that query. Across seven independently trained CIFAR-10 checkpoints at classifier-free guidance (CFG) scale 2, a longer horizon worsens FID at early steps but improves it at the midpoint. To probe whether query time alone determines this response, we fix the midpoint query and current/subsequent guidance while changing only earlier CFG. In held-out factorial outcomes on five previously studied checkpoints, the mean midpoint horizon benefit increases from 0.109 to 1.308 FID, indicating sensitivity to the incoming-state distribution. A frozen four-query rule reduces mean CIFAR-10 FID by 2.023 without retraining or extra queries, including gains on five checkpoints trained after rule selection. A milder rule also improves mean FID by 1.041 on a stronger frozen grid, with fixed-rule FID gains on the evaluated public ImageNet-256 and AFHQ checkpoints. However, matched-budget calibration favors query-start shifting in historical CIFAR tests, and the tested Horizon allocation adds no FID benefit beyond Grid+Start on two public ImageNet models. Thus, prediction horizon is a distinct, position- and state-sensitive inference control, but not a universally preferred one.