In this paper, we propose Decomposable Flow Matching (DFM), a novel framework that independently applies Flow Matching to each level of a user-defined multi-scale representation (e.g., Laplacian pyramid) to address the computational cost of high-dimensional visual modality generation. DFM overcomes the complexity of existing multi-stage generative models (requiring custom diffusion formulas, decomposition-dependent stage transitions, temporal samplers, or model cascades) and improves the visual quality of images and videos with a single model. Experimental results show that the FDD score is improved by 35.2% over the baseline architecture and by 26.4% over the state-of-the-art baseline on the Imagenet-1k 512px dataset. Furthermore, when applied to fine-tuning large-scale models such as FLUX, it shows faster convergence speed on the training distribution. This is possible with minimal modifications to the existing training pipeline.
