Towards Hierarchical Rectified Flow
Authors: Yichi Zhang, Yici Yan, Alex Schwing, Zhizhen Zhao
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically verify this on synthetic 1D and 2D data as well as MNIST, CIFAR-10, and Image Net-32 data. |
| Researcher Affiliation | Academia | 1University of Illinois Urbana-Champaign |
| Pseudocode | Yes | Algorithm 1: Hierarchical Rectified Flow Training Algorithm 2: Hierarchical Rectified Flow Sampling |
| Open Source Code | Yes | Our code is available at: https://riccizz.github.io/HRF/. |
| Open Datasets | Yes | We empirically verify this on synthetic 1D and 2D data as well as MNIST, CIFAR-10, and Image Net-32 data. |
| Dataset Splits | Yes | For each dataset in the low-dimensional experiments, we use 100,000 data points for training and another 100,000 data points for evaluation. |
| Hardware Specification | Yes | We train all models on a single NVIDIA RTX A6000 GPU. |
| Software Dependencies | No | In our experiments, we use the RK45 ODE solver (Dormand & Prince, 1980) provided by the scipy.integrate.solve_ivp package. We use atol = 1e-5 and rtol = 1e-5. |
| Experiment Setup | Yes | We use the Adam optimizer with β1 = 0.9, β2 = 0.999, and ϵ = 10^-8, with no weight decay. For MNIST, the U-Net has channel multipliers [1, 2, 2], while for CIFAR-10 and Image Net-32, the channel multipliers are [1, 2, 2, 2]. The learning rate is set to 1e-4 with a batch size 128 for MNIST and CIFAR-10. For Image Net-32, we increase the batch size to 512 and adjust the learning rate to 2e-4. We train all models on a single NVIDIA RTX A6000 GPU. For MNIST, we train both the baseline and our model for 150,000 steps while we use 400,000 steps for CIFAR-10. |