Method

Decomposable Flow Matching (DFM): A generative model combining multiscale decomposition with Flow Matching. DFM progressively synthesizes different representation scales by generating coarse-structure scale first and incrementally refining it with finer scales.

DFM Architecture: Our framework (DFM) progressively synthesizes images by combining multiscale decomposition with Flow Matching. We modify the DiT architecture to use per-scale patchification and timestep-embedding layers while keeping the core DiT architecture untouched. DFM Training: We sample the stage count from a categorical distribution, then draw each stage flow-timestep from a logit-normal distribution biased toward lower noise in early stages, and finally train one DiT backbone to jointly predict all stage-wise velocities. DFM Inference: We denoise the coarsest stage first for T1 steps and activate next stages after the previous ones reach a predetermined per-stage noise threshold. This yields previewable outputs with earlier stages build global structure while later ones generate finer details. DFM Architecture Training Inference

Results

Across image and video generation, DFM outperforms the best-performing baselines, achieving the same Fréchet DINO Distance (FDD) of Flow Matching baselines with up to 2x less training compute.

Qualitative Results

Large-Scale Finetuning: Finetuning FLUX-dev with DFM (FLUX-DFM) achieves superior results than finetuning with standard full-finetuning (DFM-FT) for the same training compute.

Training From Scratch for Image Generation: When trained from scratch on ImageNet-1k 512px, DFM achieves better quality than baselines using the same training resources.

Training From Scratch for Video Generation: DFM is also suited for video generation, achieving better structural and visual quality than baselines when trained on the Kinetics-700 dataset with the same compute budget.

Ablations: We found that DFM benefits from more sampling steps in the coarse-structure stage and needs only a few in the high-frequency stage, and it stays largely insensitive to the choice of sampling per-stage noise threshold, especially at high CFG values.

Citation

If you find this paper useful in your research, please consider citing our work:

@article{dfm,
title={Improving Progressive Generation with Decomposable Flow Matching},
author={Moayed Haji-Ali and Willi Menapace and Ivan Skorokhodov and Arpit Sahni and Sergey Tulyakov and 
  Vicente Ordonez and Aliaksandr Siarohin},
journal={arXiv preprint arXiv:2506.19839}
year={2025}}