Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation
Authors: Dogyun Park, Taehoon Lee, Minseok Joo, Hyunwoo J. Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Image Net 256 256 demonstrate that BFM establishes a substantially improved Pareto frontier over existing Flow Matching methods, achieving 2.1 to 4.9 accelerations in inference complexity at comparable generation performance. |
| Researcher Affiliation | Academia | Dogyun Park Korea University EMAIL Taehoon Lee KAIST EMAIL Minseok Joo Korea University EMAIL Hyunwoo J. Kim KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Training Blockwise Flow Matching (BFM) Algorithm 2 Training Feature Residual Network (FRN) Algorithm 3 Inference Algorithm 4 Efficient inference with Feature Residual Network |
| Open Source Code | Yes | Code is available at https://github.com/mlvlab/BFM. |
| Open Datasets | Yes | Extensive experiments on Image Net 256 256 demonstrate that BFM establishes a substantially improved Pareto frontier over existing Flow Matching methods |
| Dataset Splits | Yes | Most of the experiments are conducted on the Image Net dataset at a resolution of 256 256, unless stated otherwise. We follow the experimental setup established by Si T [22] and REPA [26]. |
| Hardware Specification | Yes | All experiments are conducted with a global batch size of 864 using eight NVIDIA A100 GPUs. |
| Software Dependencies | No | We implement our model based on the original Si T implementation [22] with recent improvements such as Swi GLU [68] activations, RMS normalization [68], and Rotary Positional Embeddings (Ro PE) [69]. Following Fi Tv2 [6], we reassign model parameters and adopt Ada LNLo RA [70] within transformer blocks to mitigate the parameter overhead of Ada LN modules. |
| Experiment Setup | Yes | Table 7: Architecture and training configurations of Blockwise Flow Matching Training Config. Optimizer Adam W Learning rate 1e-4 Batch size 864 λ 0.5 |