[Paper] One-step Latent-free Image Generation with Pixel Mean Flows

Source: arXiv - 2601.22158v1

Overview

The paper introduces Pixel MeanFlow (pMF), a novel approach that generates high‑resolution images in a single forward pass without relying on latent representations. By decoupling the network’s output space from its loss space, pMF bridges two recent trends—one‑step sampling and latent‑free generation—delivering ImageNet‑level quality (FID ≈ 2.2 @ 256², 2.5 @ 512²) with dramatically reduced inference cost.

Key Contributions

• One‑step, latent‑free generation: Produces photorealistic images in a single network evaluation, eliminating the multi‑step diffusion/flow pipeline.
• MeanFlow loss formulation: Introduces a velocity‑field‑based loss that operates on the MeanFlow of pixel intensities, while the network predicts directly on the image manifold.
• Simple image‑velocity transformation: Provides a mathematically tractable mapping between pixel values and their average velocity field, enabling stable training.
• State‑of‑the‑art FID scores on ImageNet at 256×256 and 512×512 resolutions, matching or surpassing multi‑step diffusion baselines.
• Scalable architecture: Compatible with existing convolutional and transformer backbones, making it easy to plug into current generative pipelines.

Methodology

1. Separate output & loss spaces

   • Output: The network is trained to predict the final image x (i.e., the point on the low‑dimensional image manifold).
   • Loss: Instead of a pixel‑wise L2 loss, the authors define a MeanFlow loss in the velocity space, measuring how the predicted image’s average pixel motion aligns with a ground‑truth flow field derived from the data distribution.

2. MeanFlow transformation

   • For any image x, they compute an average velocity field v = M(x) that captures the direction and magnitude of pixel changes needed to reach x from a reference distribution.
   • The inverse mapping M⁻¹(v) reconstructs an image from its velocity field, ensuring a bijective relationship that keeps training stable.

3. Training pipeline

   • Sample a random noise image z.
   • Pass z through the generator to obtain a candidate image x̂.
   • Compute v̂ = M(x̂) and compare it to the target velocity v* = M(x_real) using a simple L2 loss in velocity space.
   • Back‑propagate the loss to update the generator; no iterative refinement or latent encoder is needed.

4. Network design

   • The authors use a standard UNet‑style backbone with attention blocks, but the core idea works with any architecture that can map noise to pixel space.

Results & Findings

• Speed: Generation time drops from ~1 s (50‑step diffusion) to <10 ms on a single GPU, a >100× speedup.
• Quality: Visual inspection shows crisp textures and faithful class semantics, comparable to state‑of‑the‑art diffusion models.
• Stability: Training converges in ~300 k iterations, similar to conventional diffusion training, despite the radically different loss formulation.

Practical Implications

• Real‑time content creation: Developers can embed high‑quality image synthesis directly into interactive applications (e.g., game asset generation, UI mock‑ups) without waiting for multi‑step sampling.
• Edge deployment: The one‑step nature reduces memory bandwidth and compute cycles, making it feasible to run on consumer GPUs, mobile SoCs, or even WebGPU environments.
• Simplified pipelines: No need for separate latent encoders, scheduler designs, or sampling heuristics—just a single forward pass. This lowers engineering overhead for SaaS platforms offering on‑demand image generation.
• Cost reduction: Cloud inference costs drop dramatically when each request consumes milliseconds instead of seconds, enabling scalable APIs for generative services.
• Foundation for downstream tasks: The velocity‑field perspective could be repurposed for image editing, style transfer, or video frame interpolation, where controlling pixel motion is valuable.

Limitations & Future Work

• Training data dependence: The MeanFlow mapping is learned from the training distribution; out‑of‑distribution prompts may still suffer from mode collapse or artifacts.
• Limited conditional control: The current formulation focuses on unconditional generation; extending pMF to text‑to‑image or class‑conditional settings requires additional conditioning mechanisms.
• Theoretical guarantees: While the bijective mapping between image and velocity spaces works empirically, a rigorous analysis of its expressiveness and invertibility is left for future research.
• Broader benchmarks: Experiments are confined to ImageNet; evaluating on domain‑specific datasets (e.g., medical imaging, satellite data) will test the method’s generality.

Overall, Pixel MeanFlow marks a significant step toward ultra‑fast, high‑fidelity generative models that can be readily adopted by developers building the next generation of AI‑powered visual tools.

Authors

• Yiyang Lu
• Susie Lu
• Qiao Sun
• Hanhong Zhao
• Zhicheng Jiang
• Xianbang Wang
• Tianhong Li
• Zhengyang Geng
• Kaiming He

Paper Information

• arXiv ID: 2601.22158v1
• Categories: cs.CV
• Published: January 29, 2026
• PDF: Download PDF