Improved Training Technique for Latent Consistency Models
Authors: Minh Quan Dao, Khanh Doan, Di Liu, Trung Le, Dimitris Metaxas
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We measure the performance of our proposed technique on three datasets: Celeb A-HQ (Huang et al., 2018), FFHQ (Karras et al., 2019), and LSUN Church (Yu et al., 2015), at the same resolution of 256 x 256. ... We use two well-known metrics, Fr echet Inception Distance (FID) (Naeem et al., 2020) and Recall (Kynk a anniemi et al., 2019), for measuring the performance of the model given the training data and 50K generated images. (Section 5.1) ... We ablate our proposed techniques on the Celeb A-HQ 256 x 256 dataset, with all FID and Recall metrics measured using 1-NFE sampling. All models are trained for 1,400 epochs with the same hyperparameters. (Section 5.2) |
| Researcher Affiliation | Collaboration | Quan Dao Rutgers University EMAIL Khanh Doan Qualcomm Vietnam Company Limited EMAIL Di Liu Rutgers University EMAIL Trung Le Monash University EMAIL Dimitris Metaxas Rutgers University EMAIL |
| Pseudocode | No | The paper describes the proposed methods and techniques using prose and mathematical equations (e.g., eq. 6, 7, 8, 9, 11, 12), but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The implementation is released here: https://github.com/quandao10/s LCT/ |
| Open Datasets | Yes | We measure the performance of our proposed technique on three datasets: Celeb A-HQ (Huang et al., 2018), FFHQ (Karras et al., 2019), and LSUN Church (Yu et al., 2015) |
| Dataset Splits | No | The paper mentions using Celeb A-HQ, FFHQ, and LSUN Church datasets but does not explicitly provide details about the training, validation, and test splits used for these datasets. It only refers to generating 50K samples for evaluation. |
| Hardware Specification | No | The paper states '(*) means training on our machine with the same diffusion forward and equivalent architecture' in the context of Table 1, but does not provide any specific details about the hardware specifications (e.g., GPU models, CPU, memory) of this machine. |
| Software Dependencies | No | The paper mentions using the POT library in Section 4.4 and pretrained VAE KL-8 in Section 5.1, but does not provide specific version numbers for any software dependencies or libraries used in their implementation. |
| Experiment Setup | Yes | All models are trained for 1,400 epochs with the same hyperparameters. (Section 5.2). Additionally, the paper discusses specific training settings such as EMA decay rate, Fourier scaling factor, dropout rate, exponential curriculum for discretization steps N, discrete Lognormal distribution for timestep sampling, and the adaptive c scheduler (Sections 3, 4.5). |