Learning Anisotropic Prediction Targets for Flow Models
Read the original on arXiv Machine Learning →The Flow has not summarised this story yet — read it at arXiv Machine Learning.
The Flow has not summarised this story yet — read it at arXiv Machine Learning.
arXiv:2607. 12171v1 Announce Type: cross Abstract: In rectified-flow-based generative models, the neural network can be trained to predict two different targets, such as the instantaneous velocity or the data endpoint, to perform denoising.
In rectified-flow-based generative models, the neural network can be trained to predict two different targets, such as the instantaneous velocity or the data endpoint, to perform denoising. Although prior work shows that these parameterizations lead to different empirical behaviors, the mechanisms underlying their respective advantages remain to be underexplored, and how to combine them effectively is still unclear.
Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to be learned implicitly.
The paper revisits pixel-space diffusion models, showing that a vanilla diffusion transformer can generate directly in pixel space by swapping velocity prediction for clean image prediction. It introduces MFlow, an adaptive parameterization that interpolates between these two prediction modes based on data representation and model capacity, and demonstrates competitive performance on ImageNet, UCF‑101, and text‑to‑image tasks. Additionally, the authors add ConvRefiner, a lightweight convolutional module trained post‑hoc to restore high‑frequency details while preserving FID.
The paper introduces Wiener-informed Decoding (WiD), a method that leverages frozen Wiener estimates and statistical knowledge to improve the training of pixel-space flow models. By integrating a trainable Wiener prediction path with coefficient interfaces that inject features into the native decoder, WiD accelerates training, enhances generation quality, and yields smoother gradient trajectories across various model scales and decoder architectures. Experiments on ImageNet 256×256 show that PixelDiT-XL with WiD achieves an FID of 1.53 after 500 epochs, matching the quality of the baseline while requiring 37.5% fewer training epochs.
The paper introduces a Focal Log-Frequency Loss (f-loss) to counteract the spectral imbalance in pixel-space flow matching, where low frequencies dominate training. By balancing learning signals across frequencies and combining early frequency-domain supervision with later pixel-space refinement, the method accelerates convergence by up to 40% and improves FID and perceptual fidelity across multiple model scales. It requires no architectural changes and can replace existing flow matching losses as a drop‑in solution.