Self-Correcting Self-Consuming Loops for Generative Model Training
Authors: Nate Gillman, Michael Freeman, Daksh Aggarwal, Chia-Hong Hsu, Calvin Luo, Yonglong Tian, Chen Sun
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically validate the effectiveness of self-correcting self-consuming loops on the challenging human motion synthesis task, and observe that it successfully avoids model collapse, even when the ratio of synthetic data to real data is as high as 100%. |
| Researcher Affiliation | Collaboration | 1Brown University 2Google DeepMind. Correspondence to: Nate Gillman <nate_gillman@brown.edu>, Chen Sun <chensun@brown.edu>. |
| Pseudocode | Yes | Algorithm 1 Iterative Fine-tuning of a Generative Model With Correction |
| Open Source Code | Yes | We have released all the code associated with this paper.1 1Project page: https://nategillman.com/sc-sc.html |
| Open Datasets | Yes | We preprocess the MoVi (Ghorbani et al., 2021) subset of Human ML3D (Guo et al., 2022) using the official code implementation of Human ML3D. |
| Dataset Splits | No | The paper specifies a 'train set of size n = 2794 and a test set of size 546' and also smaller training sets, but does not explicitly mention a 'validation' split or its details. |
| Hardware Specification | No | The paper states: 'Our research was conducted using computational resources at the Center for Computation and Visualization at Brown University.' This does not provide specific hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions 'AdamW' and 'UHC' as components used, but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We experiment with synthetic augmentation percentages λ {0.05, 0.10, 0.15, 0.20, 0.25} on the larger dataset; we set the number of batches seen during generation 0 to be 3125, and the number of batches seen for each later generation to be m = 625. ... We use the same hyperparameters as those used for MDM, including batch size 64, AdamW (Loshchilov & Hutter, 2019) with learning rate 1e-4, and classifier-free guidance parameter 2.5. |