Rethinking Pixel Mean Flows via Interval Denoiser

2026-08-05

AI Summary

This work addresses the high computational cost of multi-step sampling and the reconstruction bottleneck of autoencoders in few-step generative modeling by proposing the Interval Denoiser framework. Built upon flow-matching ordinary differential equations (ODEs), the method explicitly reveals—within any arbitrary time interval—that predictions reside on a low-dimensional manifold, enabling an exact analytical mapping of intermediate trajectory states. This formulation avoids empirical algebraic substitutions and accurately isolates the pure time derivative, yielding unbiased gradients suitable for first-order optimization. Integrated with residual clipping and a temporal curriculum strategy, the framework supports end-to-end pixel-level training. Trained from scratch on ImageNet at 256×256 resolution, it achieves a single-step (1-NFE) FID of 4.55 and a two-step (2-NFE) FID of 3.98 without relying on perceptual losses.

Abstract

Modern diffusion and flow-based models are increasingly moving toward few-step, latent-free generation to bypass the computational overhead of multi-step sampling and the reconstruction bottlenecks of external autoencoders. We propose the Interval Denoiser, a theoretically rigorous framework for latent-free generation. Derived directly from the flow matching ODE, it establishes an exact analytical mapping for intermediate trajectory states. Unlike prior formulations, our prediction is shown to reside on a low-dimensional manifold across any time interval, making the regression tractable for a network operating directly on pixels. Furthermore, by avoiding empirical algebraic substitutions, our formulation correctly isolates the pure time derivative to prevent biased gradient evaluations and ensure exact first-order optimization. By analyzing this objective, we equip our framework with residual clipping and a time-sampling curriculum, enabling effective long-interval training and improving few-step performance. Trained from scratch on ImageNet 256x256, our model achieves an FID of 4.55 in one step (1-NFE) and 3.98 in two steps (2-NFE) without perceptual losses.

Problem

• latent-free generation
• few-step sampling
• flow matching
• image generation
• computational overhead

Innovation

• Interval Denoiser
• latent-free generation
• flow matching
• few-step sampling
• low-dimensional manifold