CrossFlow: One-Step Generation Across Latent and Pixel Spaces

2026-06-18

AI Summary

Traditional generative models are trained in latent space yet rely on separate decoders to synthesize pixel-space images, leading to a misalignment between training objectives and final image quality. This work proposes CrossFlow, a one-step cross-space flow matching method that eliminates the need for explicit velocity fields by directly mapping noisy latent representations to pixel-space images, thereby unifying the roles of generator and decoder. By integrating a latent encoder with pixel-level perceptual and adversarial losses, CrossFlow enables end-to-end cross-space mapping within a flow matching framework. Evaluated on ImageNet-1k at 256×256 resolution, the method achieves a state-of-the-art FID of 1.62 with only a single forward pass, significantly enhancing both generation efficiency and fidelity.

Abstract

Most diffusion and flow-matching generators define the prior, probability path, and prediction target in the same representation space. Latent diffusion improves efficiency by moving this path into an autoencoder latent space, but the final sample is still produced by a separately trained decoder. This separation creates a mismatch: the generator is optimized for latent-space prediction, while final quality depends on how the decoder handles generated latents that may differ from clean encoder outputs. We introduce CrossFlow, a cross-space flow formulation that maps noisy latent inputs directly to pixel-space images. The key technical step is a velocity-free one-step objective: the latent trajectory defines the training path, but the supervised prediction is an image rather than a latent displacement. This lets one model act both as a one-step latent-to-pixel generator and as a decoder replacement for latent diffusion pipelines. On class-conditional ImageNet-1k at $256\times256$, CrossFlow-XL achieves 1.62 FID with one function evaluation. Ablations show that the latent encoder and pixel-space perceptual and adversarial losses are important for fidelity. These results indicate that cross-space flow objectives can combine the efficiency of latent representations with direct pixel-space supervision, without requiring a separate decoder at inference.

Problem

• latent space
• pixel space
• decoder mismatch
• generative modeling
• flow matching

Innovation

• cross-space flow
• one-step generation
• latent-to-pixel mapping
• decoder-free generation
• flow matching