Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlow
MeanFlow predicts interval-average velocities, but conventional few-step sampling ties the interval queried from the model to the duration of the state update.
ProofPaper ↗
Key points
- 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.
Sources (1)
- [1]Rethinking Prediction Horizons: Decoupling Model Queries from State Updates in Few-Step MeanFlowarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 12:11 PM
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.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026Data-free Multistep Diffusion Distillation
- Oct 13, 2026Diffusion without CFG? Correction Training for Diffusion Models
- Oct 13, 2026Dog-Breed Text-to-image Flow Matching from Pretraining to Fine-Tuning on 4 Consumer GPUs
- Oct 13, 2026LoopREG: Representation Supervision Enables Test-Time Recurrent Depth in Diffusion Transformers
- Oct 13, 2026Efficient Scene-Text Refinement via Pixel-Space Grouped Image Generation
- Oct 12, 2026Bottlenecked semantic compression by representation reconstruction