emerging
CrossFlow: One-Step Generation Across Latent and Pixel Spaces
· Peking University, Tencent, Fudan University · 2026-06 · arXiv:2606.19970
Why it matters
A one-step flow from noisy latents to pixel-space images, FID 1.62 at a single function evaluation. A latent-to-pixel bridge rather than tokenizer-free generation, recorded here so the distinction stays visible.
What this paper does
read: full textBefore this
By 2026, tokenizer-free pixel diffusion (pixelflow-2025, sid2-2024, hdit-2024) argued generation should skip latents altogether, while latent-forcing-2026 argued latents remain useful as an intermediate scratchpad within an otherwise pixel-output generator; conventional latent diffusion still optimized its generator purely for latent-space prediction and handed off to a separately trained decoder network that was never exposed to the generator's own possibly imperfect outputs during its training, creating a train and inference mismatch at the latent-to-pixel handoff.
The problem
How to remove the mismatch between a latent diffusion generator, optimized for latent-space prediction, and its decoder, trained independently and never on the generator's own outputs, while keeping the efficiency of running the diffusion or flow trajectory in a compressed latent space.
The idea
Keep the noisy trajectory in latent space, as ordinary latent diffusion does, but change what the model is trained to predict: instead of a latent-space velocity or displacement, have the network directly predict the clean pixel-space image at every point along the latent trajectory, so a single model can act as both a one-step latent-to-pixel generator from noise and as a drop-in replacement for a separately trained decoder inside an existing latent diffusion pipeline.
How it works
CrossFlow defines a cross-space flow: the prior, the probability path, and the intermediate noisy states are all in the autoencoder's latent space, as in ordinary latent flow matching, but the training target at every point on that path is the clean pixel-space image, not a latent displacement vector, described as a velocity-free one-step objective. Because the network's output head is bound to predict a clean image directly rather than a velocity, no multi-step ODE integration is needed at inference; a single forward pass maps a noisy latent straight to a pixel-space image. This training scheme lets the same trained network be used two ways: as a genuine one-step generator, sampling a noisy latent and mapping it to pixels in one function evaluation, and as a decoder substitute inside conventional multi-step latent diffusion pipelines, since it consumes latents, including imperfect ones produced by another generator, and emits images the way a decoder would.
Evidence
On class-conditional ImageNet-1k at 256x256, CrossFlow-XL (461M parameters) reaches FID 1.62 with a single function evaluation, one forward pass with no multi-step sampling. Ablations on latent encoder choice, evaluated on a smaller CrossFlow-B model, show the default VA-VAE encoder giving 2.54 FID versus 3.81 FID with a FLUX.2 encoder and 4.08 FID with an SD-VAE encoder, so encoder choice matters substantially. Loss ablations on CrossFlow-B show removing the GAN loss degrades FID from 2.54 to 8.86, removing the perceptual loss degrades it to 30.21, removing both degrades it to 50.65, and a further variant setting phi(r) to zero gives 8.70, together showing the pixel-space perceptual and adversarial losses are load-bearing for fidelity rather than optional polish.
Limitations
The authors concede that training the encoder and the cross-space generator simultaneously exhibits optimization instabilities in practice, so the reported system does not use single-stage end-to-end training: the encoder is trained first and then frozen while the generator is trained, with the authors explicitly deferring stable single-stage joint training strategies to future work. The clean single-model story therefore holds only at inference time, not at training time.
Why it matters
On the atlas's question, whether CrossFlow goes from noisy latents to pixel-space images in one step, the fetched text confirms this directly, and the paper explicitly frames its contribution as fixing a latent-diffusion-specific problem, generator and decoder mismatch, through cross-space supervision, not as eliminating the latent or arguing against having one. It should be recorded as a latent-to-pixel bridge and possible decoder replacement, distinct in kind from tokenizer-free generation, which removes the latent from the pipeline entirely, and also distinct from latent-forcing-2026's scratchpad framing, which keeps pixels as the model's own multi-step output with the latent feeding pixel denoising rather than being mapped to pixels by a one-step head; it occupies its own third position in this space.
Abstract, in the authors' own words
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.
Research lines
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.
This is a emerging 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.