Research map / Image generation / Fast diffusion sampling

Asymmetric Flow Models

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Rectified flow straightening · 0 citations (OpenAlex)

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 AsymFlow, a rank-asymmetric velocity parameterization (low-rank noise, full-rank data) that enables pixel-space flow models without architecture changes, and a latent-to-pixel finetuning route from FLUX.2 klein 9B.

Problem

Flow-based generation in high-dimensional pixel space is hard because velocity prediction requires modeling high-dimensional noise; existing latent flow models cannot be easily converted into pixel generators.

Approach

Predict an asymmetric velocity P·eps − x0 where the noise term is restricted to a low-rank patch subspace while data stays full-dimensional; the full-rank velocity is recovered analytically. A Procrustes-aligned linear lift lets a latent flow model be reinterpreted as a low-rank pixel flow for finetuning.

Key results

• AsymFlow-H/16 with REPA reaches 1.57 FID on ImageNet 256x256 (1.76 without REPA), versus JiT-H/16 at 1.86.
• AsymFLUX.2 klein (from FLUX.2 klein Base 9B) reaches HPSv3 10.66, DPG 86.8, GenEval 0.82, improving over its latent base (9.50, 85.2, 0.80) and PixelDiT-T2I (8.95, 83.5, 0.74).
• Without sigma_min clamping, JiT degrades by 1.37 FID while AsymFlow degrades by only 0.52 FID (r=8).

Surprising findings

• Plain REPA improves AsymFlow from 1.76 to 1.57 FID while it only improves JiT-H/16 from 1.86 to 1.81 with PixelREPA's additional designs.
• A DDT-head pixel finetuning baseline with more parameters falls behind the simpler AsymFlow finetuning on all text-to-image metrics.

Limitations

• Latent-to-pixel finetuning assumes a good patch-level linear lift and may fail when the latent space does not preserve pixel structure (e.g., RAE models).
• Hyperparameters for perceptual correction were not systematically swept due to compute constraints.
• Perceptual/variance-reduction term introduces excessive noise at low noise levels without the LPIPS correction.

Experimental setup

Datasets: ImageNet LAION-Aesthetics COCO 2014 validation

Benchmarks: ImageNet 256x256 HPSv3 DPG-Bench GenEval COCO-10K

Metrics: FID IS HPSv3 HPSv2.1 VQAScore CLIP score pFID DPG GenEval

Baselines: JiT-H/16 JiT-G/16 PixelDiT-T2I PixelDiT-XL/16 DeCo-XL/16 PixelREPA-H/16 PixelGen-XL/16 DiP-XL/16 PixNerd-XL/16 FLUX.2 klein Base FLUX.1 dev Qwen-Image

Builds on: JiT REPA flow matching DDT FLUX.2 klein

Closest papers by meaning

• Cross Flow: One-Step Generation Across Latent and Pixel Spaces (NeurIPS 2026 · similarity 0.88)
• Improving Progressive Generation with Decomposable Flow Matching (NeurIPS 2025 · similarity 0.88)
• FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models (CVPR 2024 · similarity 0.87)
• Balancing Frequencies and Pixels in Flow Matching (NeurIPS 2026 · similarity 0.87)
• PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.86)
• Taming generative video models for zero-shot optical flow extraction (NeurIPS 2025 · similarity 0.86)