Research map / Image generation / Fast diffusion sampling
Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation
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 Blockwise Flow Matching, which splits the flow trajectory into temporal segments each handled by a smaller specialized velocity block, plus semantic feature guidance (DINOv2-aligned features) and a lightweight residual feature approximator to cut inference FLOPs on ImageNet 256.
Problem
Single large flow matching networks must handle very different signal characteristics across timesteps and incur high inference cost because the full model is evaluated at every solver step.
Approach
Partition the generative trajectory into M temporal segments, each modeled by its own smaller velocity block so only one block runs per step; condition blocks on semantic features from a shared alignment network, and approximate those features within each segment via a residual network to avoid repeated evaluation.
Key results
- BFM-XL with SemFeat reaches FID 1.75 on ImageNet 256 with classifier-free guidance, vs. 1.80 for REPA-XL, at 107.8 GFLOPs vs. 114.5.
- BFM-XL with SemFeat and residual approximation reaches 37.8 GFLOPs (vs. 114.5 for REPA-XL) at FID 2.03, reported as 2.1x to 4.9x inference-complexity acceleration.
- With six solver steps, BFM-XL_SF achieves FID 8.01 vs. 12.91 for SiT-XL and 13.02 for REPA-XL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Partition the flow into 6 specialized segments | SiT-S flow matching · ImageNet 256×256 | FID: 82.6 -> 81.5 | Helped |
| Add Semantic Feature Guidance (ablation) | BFM-S · ImageNet 256×256 | FID: 81.5 -> 66.9 | Helped |
| Add Feature Residual Approximation (ablation) | BFM-S-SF · ImageNet 256×256 | GFLOPs: 5.01 to 2.96 | Helped |
| Add Feature Residual Approximation (ablation) | BFM-S-SF · ImageNet 256×256 | FID: 66.9 -> 68.3 | Hurt |
| Use BFM-XL with Semantic Feature Guidance instead of REPA-XL | ImageNet 256×256 generation · ImageNet 256×256 | FID: 1.80 -> 1.75 | Helped |
| Use Feature Residual Approximation | BFM-XL-SF · ImageNet 256×256 | GFLOPs: 114.5 to 37.8 | Helped |
| Use 6 solver steps rather than 246 | BFM-XL-SF · ImageNet 256×256, 6 solver steps | FID: 1.36 -> 8.01 | Hurt |
| Use BFM-SF with MeanFlow | MeanFlow · ImageNet 256×256, 4 function evaluations | FID: 13.2 -> 9.7 | Helped |
| Use BFM-S/4 instead of SiT-S/4 | ImageNet 512² generation · ImageNet 512² | FID: 102.54 -> 87.85 | Helped |
| Increase segment count from 4 to 6, 8, or 12 while keeping layers per segment fixed (ablation) | BFM · ImageNet 256×256, fixed layers per segment | FID: 85.1 -> 81.5 -> 79.1 -> 76.7 | Helped |
| Increase segment count beyond 6 while reducing layers per segment at fixed total capacity (ablation) | BFM · ImageNet 256×256, fixed total network capacity | FID: 81.5 -> 88.1 -> 95.2 | Hurt |
| Replace REPA with SemFeat in standard FM (ablation) | FM · ImageNet 256×256 | FID: 72.9 -> 68.5 | Helped |
Surprising findings
- Splitting a model into more segments at fixed per-segment depth improves FID without adding inference FLOPs (FID 85.1 at 4 segments to 76.7 at 12 segments).
Limitations
- Relies on external pretrained encoder (DINOv2) for semantic guidance, limiting generality to domains without such encoders.
- Evaluated only on class-conditional ImageNet image generation; not validated on text-to-image or video.
- Uses uniform temporal partitioning; non-uniform partitioning is left to future work.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet 256x256 ImageNet 512x512
Metrics: FID sFID IS Precision Recall GFLOPs wall-clock runtime
Baselines: SiT-XL REPA-XL DiT-XL FiTv2-XL DyDiT-XL ADM LDM-4 U-ViT-H MeanFlow SiT-S
Builds on: Flow Matching SiT REPA FiTv2 MeanFlow
Closest papers by meaning
- Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.90)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.90)
- Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling (ICLR 2025 · similarity 0.89)
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.89)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.89)
- On the Guidance of Flow Matching (ICML 2025 · similarity 0.88)
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.