CrossFlow: One-Step Generation Across Latent and Pixel Spaces

2026-06-18 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Advanced, quick

Summary

CrossFlow introduces a novel cross-space flow formulation for image generation, addressing the efficiency-quality mismatch in latent diffusion models. Traditional latent diffusion optimizes generators in latent space, but relies on a separately trained decoder for final pixel-space output, which may struggle with generated latents. CrossFlow directly maps noisy latent inputs to pixel-space images using a velocity-free one-step objective, where the latent trajectory guides training but the prediction target is an image. This allows a single model to function as both a one-step latent-to-pixel generator and a decoder replacement, eliminating the need for a separate decoder during inference. CrossFlow-XL achieved a 1.62 FID on class-conditional ImageNet-1k at 256x256 with just one function evaluation. Its fidelity relies on a latent encoder and pixel-space perceptual and adversarial losses.

Key takeaway

For Machine Learning Engineers optimizing image generation pipelines, CrossFlow presents a significant advancement by unifying latent and pixel space generation. If your current latent diffusion models suffer from decoder-induced quality mismatches or require multiple inference steps, you should investigate cross-space flow objectives. This approach allows a single model to generate high-fidelity images directly from latents with one function evaluation, potentially streamlining your inference process and improving output quality.

Key insights

CrossFlow directly maps noisy latents to pixel-space images, combining latent efficiency with direct pixel supervision in one model.

Principles

• Cross-space flow objectives enhance efficiency.
• Direct pixel-space supervision improves fidelity.
• Integrating generator and decoder streamlines inference.

Method

CrossFlow uses a velocity-free one-step objective, defining the training path via latent trajectory while supervising prediction directly in pixel space.

In practice

• Replace separate decoders in latent diffusion.
• Generate high-fidelity images with one evaluation.
• Optimize for latent efficiency and pixel quality.

Topics

• CrossFlow
• Image Generation
• Latent Diffusion
• Flow-matching
• Pixel Space
• FID Score

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Consistency-Preserving Diverse Video Generation — A new flow-matching framework improves video diversity and temporal consistency using latent-space gradient regulation.
• Normalizing Flows with Iterative Denoising — iTARFlow advances Normalizing Flows with a likelihood-based objective and diffusion-inspired iterative denoising for image generation.
• Google's DiffusionGemma generates 256 tokens in parallel and self-corrects as it goes — DiffusionGemma applies image diffusion to text, enabling parallel, self-correcting block generation for faster inference in specific contexts.
• Aligning Latent Geometry for Spherical Flow Matching in Image Generation — Spherical Flow Matching improves image generation by aligning latent geometry on a sphere using geodesic paths.
• End-to-End Training for Unified Tokenization and Latent Denoising — Unified tokenization and latent diffusion can be jointly optimized in a single-stage training process.
• Understanding Stable Diffusion and ControlNet for a Bar Conversation — Latent space diffusion and conditional control enhance image generation efficiency and precision in models like Stable Diffusion.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.