Research map / Image generation / Fast diffusion sampling

Cross Flow: One-Step Generation Across Latent and Pixel Spaces

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Flow matching distillation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CrossFlow, a velocity-free one-step objective that maps noisy VAE-latent inputs directly to pixel images, unifying one-step generation with latent-to-pixel decoding, and reports 1.62 FID on ImageNet 256 with one NFE.

Problem

Latent diffusion models use a separately trained decoder and a generator trained only in latent space, creating a mismatch and preventing direct pixel-level supervision; standard flow/one-step methods assume input and output live in the same space.

Approach

Derives a cross-space flow objective in which the latent-space velocity is cancelled by a structural coefficient constraint, so a ViT takes noisy latents (r-conditioned) and directly predicts a pixel image, trained with L1 plus perceptual and hinge-GAN losses on a frozen VA-VAE encoder.

Key results

• CrossFlow-XL achieves 1.62 FID on ImageNet 256x256 with one function evaluation, versus iMF-XL/2 at 1.72 FID in latent space.
• CrossFlow-B/L/XL reach 2.54, 1.69, 1.62 FID with 95M, 315M, 461M parameters.
• Removing GAN loss raises FID from 2.54 to 8.86; removing perceptual loss raises it to 30.21; removing both gives 50.65 (CrossFlow-B ablation).

Surprising findings

• Removing both perceptual and GAN losses degrades FID to 50.65, showing auxiliary pixel losses are critical rather than incidental.
• Using the VA-VAE encoder beats the higher-capacity 128-channel FLUX.2 encoder (2.54 vs 3.81 FID), suggesting latent structure matters more than channel count.

Limitations

• Evaluated only on class-conditional ImageNet 256x256; no text-to-image or higher resolution results.
• Depends on a fixed pretrained encoder; end-to-end joint training was unstable in practice.
• Adversarial loss requires time-dependent weighting to avoid mode collapse at low time values.

Experimental setup

Datasets: ImageNet-1k

Benchmarks: ImageNet-1k 256x256 class-conditional generation

Metrics: FID (50k samples)

Baselines: DiT-XL/2 SiT-XL/2 SiT-XL/2 + REPA iMF-XL/2 MeanFlow-XL/2 Shortcut-XL/2 iCT-XL/2 pMF PixelDiT-XL/16 BigGAN-deep StyleGAN-XL GigaGAN

Builds on: MeanFlow Improved MeanFlow consistency models VA-VAE RAE

Closest papers by meaning

• Flowing from Words to Pixels: A Noise-Free Framework for Cross-Modality Evolution (CVPR 2025 · similarity 0.91)
• Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.89)
• Asymmetric Flow Models (NeurIPS 2026 · similarity 0.88)
• Solving Inverse Problems with FLAIR (NeurIPS 2025 · similarity 0.88)
• Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.88)
• Latent Refinement via Flow Matching for Training-free Linear Inverse Problem Solving (NeurIPS 2025 · similarity 0.88)