Easing Training Process of Rectified Flow Models Via Lengthening Inter-Path Distance

Authors: Shifeng Xu, Yanzhu Liu, Adams Kong

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on image and latent spaces of rectified flow models and diffusion models, which demonstrate DANSM s outstanding capabilities for speeding up the training process.
Researcher Affiliation Academia 1College of Computing and Data Science, Nanyang Technological University, Singapore EMAIL, EMAIL 2Institute for Infocomm Research (I2R) & Centre for Frontier AI Research, A*STAR, Singapore liu EMAIL
Pseudocode Yes Algorithm 1 Surrogate method of DANSM Input: a noise set Z of size n, and a sample set X of size n. Output: a noise-sample path set P of size n. 1: Initialize path set P = ϕ 2: for each sample x in X do 3: From Z, find z who has shortest path length with x search all in Z 4: Remove noise z from Z each noise is used only once 5: Add path p(z, x) into P 6: end for
Open Source Code Yes Code: https://github.com/shifengxu/DANSM.
Open Datasets Yes For image space, three datasets CIFAR-10 (Krizhevsky et al., 2009), Image Net64 (Deng et al., 2009), and LSUN Bedroom (Yu et al., 2015) are tested.
Dataset Splits No The paper uses standard datasets like CIFAR-10, ImageNet64, LSUN Bedroom, and FFHQ, but does not explicitly provide specific details on how these datasets were split into training, validation, and test sets. For example, Table 2 and 3 list '50K CIFAR-10 images' and '50K Bedroom latent variables' but do not provide split percentages or sample counts for each partition.
Hardware Specification No The paper does not provide specific details on the hardware used for running experiments, such as GPU models, CPU specifications, or memory.
Software Dependencies No The paper mentions the 'Scipy (Virtanen et al., 2020) library' in the context of Immiscible Diffusion's implementation, but it does not specify version numbers for Scipy or any other software dependencies used in their own methodology (e.g., deep learning frameworks, CUDA).
Experiment Setup Yes In Section 5, the paper provides several experimental setup details, including 'train the models from scratch for 500 epochs', the 'match-size' hyperparameter, normalization of 'image pixel values... to the range of [ -1, 1]', and 'latent variables are kept as their original values without scaling'. It also specifies '10-step sampling' for evaluation.