Research map / Image generation / Fast diffusion sampling
Improving Progressive Generation with Decomposable Flow Matching
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 Decomposable Flow Matching (DFM), which applies flow matching independently per level of a Laplacian pyramid with per-stage timesteps, enabling progressive image and video generation within one model.
Problem
Progressive (coarse-to-fine) generation of high-dimensional visual media usually requires multi-stage architectures, custom diffusion formulations, special samplers, or model cascades, which add complexity.
Approach
Extend Flow Matching with independent per-stage flow timesteps over a user-defined multiscale (Laplacian) decomposition, training a single shared DiT with per-scale patchification and timestep embedders, and masking loss for not-yet-active stages; sampling activates stages sequentially once a threshold is reached.
Key results
- On ImageNet-1K 512px (DiT-XL, cfg=1.5), DFM reaches FDD50K 85.8 vs 132.5 for Flow Matching and 116.2 for Pyramidal Flow.
- Finetuning FLUX with DFM (FLUX-DFM) yields FID10K 8.83 vs 12.56 for standard full finetuning (FLUX-FT) under the same compute.
- Abstract reports 35.2% FDD improvement over base architecture and 26.4% over best baseline on ImageNet 512px under the same training compute.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DFM instead of Pyramidal Flow | DiT-XL · ImageNet-1K 512px | FDD50K: 261.2 -> 200.6 | Helped |
| Use DFM instead of Pyramidal Flow | DiT-XL · ImageNet-1K 1024px | FDD10K: 351.1 -> 277.3 | Helped |
| Use DFM instead of Pyramidal Flow | DiT-XL · Kinetics-700 512px | FDD10K: 353.8 -> 336.7 | Helped |
| Apply DFM instead of standard full finetuning | FLUX-dev · Internal image validation split | FID10K: 12.56 -> 8.83 | Helped |
| Apply DFM instead of standard full finetuning | FLUX-dev · Internal image validation split | CLIPSIM: 0.3261 -> 0.3381 | Helped |
| Increase first-stage sampling probability from 0.7 to 0.95 (ablation) | DFM · ImageNet-1K 512px | FID10K: 36.02 -> 27.5 | Helped |
| Tie the timesteps of every stage (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 55.49 | Hurt |
| Remove input masking (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 29.9 | Hurt |
| Remove scale standardization (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 27.13 | Helped |
| Add stage-specific specialization to all model parameters (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 26.63 | Helped |
| Increase training batch size (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 31.95 | Hurt |
| Use smaller patch size (ablation) | DFM · ImageNet-1K 512px | FID10K: 29.56 -> 29.56 | No clear effect |
Surprising findings
- Removing standardization of per-stage inputs improved performance for the FLUX autoencoder image setting, contrary to the default.
- DiT operating in DCT frequency space performed worse than spatial-domain input, so DCT decomposition was abandoned.
Limitations
- Introduces additional training and sampling hyperparameters (stage probabilities, noise location, threshold, steps per stage).
- Artifacts such as flattened appearance in high-frequency regions like vegetation or fur.
- Relies on scale-equivariant autoencoder fine-tuning for best spectral structure.
Experimental setup
Datasets: ImageNet-1K Kinetics-700 internal image dataset
Benchmarks: ImageNet-1K 512px ImageNet-1K 1024px Kinetics-700 512px
Metrics: FID FDD IS FVD CLIPSIM
Baselines: Flow Matching Cascaded diffusion Pyramidal Flow FLUX-FT
Builds on: Flow Matching DiT Pyramidal Flow FLUX
Closest papers by meaning
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.91)
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.89)
- ProReflow: Progressive Reflow with Decomposed Velocity (CVPR 2025 · similarity 0.89)
- Cross Flow: One-Step Generation Across Latent and Pixel Spaces (NeurIPS 2026 · similarity 0.89)
- ViewFusion: Towards Multi-View Consistency via Interpolated Denoising (CVPR 2024 · similarity 0.89)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.