Consistent Diffusion Models: Mitigating Sampling Drift by Learning to be Consistent
Authors: Giannis Daras, Yuval Dagan, Alex Dimakis, Constantinos Daskalakis
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we show that enforcing CP improves the generation quality for conditional and unconditional generation on CIFAR-10, and in AFHQ and FFHQ. |
| Researcher Affiliation | Academia | Giannis Daras Department of Computer Science University of Texas at Austin Yuval Dagan Electrical Engineering and Computer Science University of California, Berkeley Alexandros G. Dimakis Department of ECE University of Texas at Austin Constantinos Daskalakis Electrical Engineering and Computer Science Massachusetts Institute of Technology |
| Pseudocode | Yes | Algorithm 1 Consistent Diffusion Models (CDM) Training |
| Open Source Code | Yes | We open-source our code and models: https://github.com/giannisdaras/cdm. |
| Open Datasets | Yes | For all our experiments, we rely on the official open-sourced code and the training and evaluation hyper-parameters from the paper Elucidating the Design Space of Diffusion-Based Generative Models [28] that, to the best of our knowledge, holds the current state-of-the-art on conditional generation on CIFAR-10 and unconditional generation on CIFAR-10, AFHQ (64 64 resolution), FFHQ (64 64 resolution). |
| Dataset Splits | Yes | We keep checkpoints during training and we report FID for 30k, 70k, 100k, 150k, 180k and 200k iterations in Table 1. We also report the best FID found for each model, after evaluating checkpoints every 5k iterations (i.e. we evaluate 40 models spanning 200k steps of training). |
| Hardware Specification | Yes | We train all our models on a DGX server with 8 A100 GPUs with 80GBs of memory each. |
| Software Dependencies | No | The paper mentions relying on 'official open-sourced code' and refers to another paper for hyperparameters, but does not explicitly list specific software packages with version numbers used for the experiments. |
| Experiment Setup | Yes | For the re-trained models on CIFAR-10, we use exactly the same training hyperparameters as in Karras et al. [28]...All models were trained for 200k iterations... For AFHQ, we dropped the batch size from the suggested value of 512 to 256... Finally, we retrain a baseline model on FFHQ for 150k iterations and we finetune it for 5k steps... We fix the number of sampling steps to 6... |