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..
Improving Progressive Generation with Decomposable Flow Matching
Authors: Moayed Haji-Ali, Willi Menapace, Ivan Skorokhodov, Arpit Sahni, Sergey Tulyakov, Vicente Ordonez, Aliaksandr Siarohin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a thorough analysis of the design space for DFM, providing insights into its behavior across different training and sampling strategies, and model architectures. Based on these findings, we instantiate DFM and evaluate its performance on established image and video generation benchmarks, specifically Image Net-1K [5] and Kinetics-700 [3]. Our results demonstrate uniform improvements over Flow Matching [27], cascaded models, and Pyramidal Flow [18]. |
| Researcher Affiliation | Collaboration | Moayed Haji-Ali , ,1,2 Willi Menapace ,2 Ivan Skorokhodov2 Arpit Sahni2 Sergey Tulyakov2 Vicente Ordonez1 Aliaksandr Siarohin2 1Rice University 2Snap Inc |
| Pseudocode | Yes | Algorithm 1 Sampling procedure of training timesteps across multi-scale stages |
| Open Source Code | No | At the time of submission we do not possess authorization to publish our code. However, we comprehensively describe our framework and evaluation procedures to allow reproducibility and produce our main results on public datasets. |
| Open Datasets | Yes | Training details We train our models on the Image Net-1K [5] and Kinetics-700 [3] datasets. Our main image experiments are trained and evaluated on Image Net-1K [5] which has a research-only, non-commercial license. Our video experiments are trained and evaluated on Kinetics-700 [3], which is available under the Creative Commons Attribution 4.0 (CC BY 4.0) license. |
| Dataset Splits | No | The paper mentions using Image Net-1K and Kinetics-700 datasets and specifies evaluation samples (e.g., "50k generated samples"), but it does not explicitly state the train/test/validation splits for these datasets (e.g., percentages, sample counts, or clear reference to standard splits used for training and testing). |
| Hardware Specification | Yes | Image Net-1K 512px experiments are trained on a single node containing 8 H100 GPUs, whereas Image Net-1K 1024px and Kinetics-700 512px experiments used 2 nodes of the same type. Ablations were trained on a single node. |
| Software Dependencies | No | The paper mentions the use of 'Adam optimizer' and 'Di T architecture', and 'Euler ODE sampler', but it does not provide specific version numbers for any software libraries, frameworks (like PyTorch or TensorFlow), or programming languages used in the implementation. |
| Experiment Setup | Yes | We train using the Adam optimizer (β1 = 0.9, β2 = 0.99, ϵ = 10 8) and a base learning rate of 0.0001 with 10k linear warmup steps, weight decay of 0.01, and total batchsize of 256. We train ablations on Image Net-1K for 600k steps, and the main experiment for 500k and 350k steps, respectively, for 512px and 1024px resolution. For Kinetics-700, we train the main experiments for 200k steps. |