strong-followup
One-step Latent-free Image Generation with Pixel Mean Flows
Tianhong Li, Zhengyang Geng, Kaiming He · MIT · 2026-01 · arXiv:2601.22158
Why it matters
Carries MeanFlow into pixel space, removing the tokenizer entirely while keeping one-step generation. The clearest convergence point of the pixel-space and one-step lines: if it holds, neither a VAE nor a foundation encoder is needed at either training or sampling time.
What this paper does
read: full textBefore this
MeanFlow methods such as iMF-XL/2 already achieved one-step generation with FID 1.72 on ImageNet 256x256, but did so in VAE latent space, still requiring a pretrained tokenizer and its decoder at both training and inference. Separately, pixel-space diffusion work such as JiT showed a plain transformer can generate directly on raw pixels by predicting clean data (x-prediction) rather than noise, under the manifold hypothesis that denoised images lie on a low-dimensional manifold even though noised quantities do not, but JiT still needed many sampling steps.
The problem
One-step generation and latent-free pixel-space generation had each been solved separately, but combining them poses a direct conflict. One-step MeanFlow training requires the network to model trajectories across arbitrary (r, t) time pairs, a task suited to a velocity-space target, while pixel-space generation requires an x-prediction target aligned to the low-dimensional image manifold, since a network asked to predict a noised quantity in raw pixel space rather than a denoised one fails catastrophically. No method had achieved genuine one-step, latent-free, pixel-space generation.
The idea
Separate what the network outputs from what the loss is computed on. Let the network predict clean pixels directly, on the presumed low-dimensional image manifold, while defining the training loss through MeanFlow's average-velocity formulation in velocity space, connecting the two through an explicit algebraic transformation between the manifold and the average-velocity field.
How it works
pMF operates entirely in pixel space with no VAE and no pretrained tokenizer at either training or sampling time. A DiT-style Vision Transformer with patch size 16 at 256x256 (patch dimension 768) or patch size 32 at 512x512 (patch dimension 3072) directly outputs a denoised-image prediction x_theta(z_t, r, t), the network's actual output space, rather than noise or velocity. This is converted to an average velocity via u_theta = (z_t - x_theta) / t, and the MeanFlow loss is computed in that velocity space following the iMF formulation V_theta = u_theta + (t - r) * JVP_stopgrad, with loss equal to the squared difference between V_theta and the instantaneous velocity v = noise - x. The JVP that computes du/dt is still used, via automatic differentiation, to build this velocity-space loss target; what changes relative to prior MeanFlow work is that the network's own parameterized output is x-prediction on the manifold rather than a raw velocity or noise prediction. Models range from B/16 (119M parameters) to H/16 (956M), trained for 160 to 360 epochs at batch size 1024 with the Muon optimizer and a perceptual loss (LPIPS plus ConvNeXt features) applied only below a noise threshold.
Evidence
pMF-H/16 at 360 epochs reaches FID 2.22 at 256x256 and pMF-H/32 reaches FID 2.48 at 512x512, both at true one-step (1 NFE) sampling with classifier-free guidance. This trails the latent-space one-step iMF-XL/2 at FID 1.72 (1 NFE) and the multi-step latent SiT-XL/2+REPA at FID 1.42 (250x2 NFE), but improves on the multi-step pixel-space JiT-G/16 at FID 1.82 (100x2 NFE) despite using far fewer function evaluations, and clearly beats the one-step pixel-space EPG-L/16 (FID 8.82, which relies on pretraining) and StyleGAN-XL (FID 2.30, at 1574 GFLOPs versus pMF-H/16's 271 GFLOPs). Ablations show the choice of prediction target is decisive. At 256x256, x-prediction reaches FID 3.53 while u-prediction (raw velocity) collapses to FID 164.89 under the same setup. Restricting the MeanFlow time sampling to only r=t or only r=0, instead of the full (r, t) plane, degrades FID to between 106 and 389.
Limitations
pMF still trails the best latent-space one-step method by about 0.5 FID points (2.22 versus 1.72), so the paper does not claim latent-free pixel-space generation is yet better than latent one-step generation, only competitive. The perceptual loss can only be applied when the added noise is below a certain threshold. The paper trains from scratch with no distillation from a larger model, and the manifold hypothesis is only empirically supported for the general r in (0, t) case, not theoretically proven; the authors state their simulations merely "suggest" the predicted x behaves like a denoised image.
Why it matters
It is the first method to combine one-step sampling with fully latent-free pixel-space generation, showing that separating the network's prediction target (x on the manifold) from the loss's computation space (MeanFlow velocity) resolves what looked like a direct conflict between the two lines of work. If the approach holds up at scale, it removes both the VAE and any foundation encoder from the generation pipeline entirely, at training and sampling time alike.
Abstract, in the authors' own words
Modern diffusion/flow-based models for image generation typically exhibit two core characteristics: (i) using multi-step sampling, and (ii) operating in a latent space. Recent advances have made encouraging progress on each aspect individually, paving the way toward one-step diffusion/flow without latents. In this work, we take a further step towards this goal and propose "pixel MeanFlow" (pMF). Our core guideline is to formulate the network output space and the loss space separately. The network target is designed to be on a presumed low-dimensional image manifold (i.e., x-prediction), while the loss is defined via MeanFlow in the velocity space. We introduce a simple transformation between the image manifold and the average velocity field. In experiments, pMF achieves strong results for one-step latent-free generation on ImageNet at 256x256 resolution (2.22 FID) and 512x512 resolution (2.48 FID), filling a key missing piece in this regime. We hope that our study will further advance the boundaries of diffusion/flow-based generative models.
Research lines
Design-axis choices
Representation
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a strong-followup entry — a short-form summary. Full paper-page explanations (before/problem/ core idea/representation/architecture/objective/conditioning/training/inference/results/ ablations/limitations) are written for landmark and core papers first; see PROJECT_STATE.md for the schedule.