CrossFlow: One-Step Generation Across Latent and Pixel Spaces

Modern latent diffusion models (LDMs) decouple image generation into two distinct phases: iterative denoising in a compressed latent space followed by a separate decoding step to reconstruct pixels. This separation creates a mismatch where the generator is optimized for latent-space prediction, but final image quality depends on how the decoder handles potentially distribution-shifted latents. The authors ask whether it is possible to unify these steps by mapping noisy latents directly to pixels in a single function evaluation, thereby enabling direct pixel-space supervision during training.

CrossFlow introduces a cross-space flow formulation that bypasses the need for latent-space velocity prediction. The key technical innovation is a velocity-free one-step objective: while the latent trajectory defines the training path, the supervised prediction is an image rather than a latent displacement. This allows the model to be trained using pixel-level perceptual and adversarial losses, which are typically unavailable to standard latent-space generators. The authors implement this using a Vision Transformer (ViT) backbone and a two-stage training process, where a pretrained image encoder is frozen, and the CrossFlow generator is trained from scratch to map latent noise to pixels.

CrossFlow-XL achieves an FID of 1.62 on class-conditional ImageNet-1k at 256x256 resolution using only a single function evaluation (1-NFE). This performance is competitive with multi-step latent diffusion baselines that require hundreds of sampling steps. The authors demonstrate that the model scales effectively with parameter count and that the inclusion of pixel-space auxiliary losses—specifically perceptual and adversarial losses—is critical for achieving high visual fidelity. Furthermore, the model can serve as a drop-in replacement for a VAE decoder in existing latent diffusion pipelines.