Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment
Authors: Yiheng Li, Heyang Jiang, Akio Kodaira, Masayoshi TOMIZUKA, Kurt Keutzer, Chenfeng Xu
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments demonstrate that our method can achieve up to 3x faster training for unconditional Consistency Models on the CIFAR dataset, as well as for DDIM and Stable Diffusion on Celeb A and Image Net dataset, and in class-conditional training and fine-tuning. |
| Researcher Affiliation | Collaboration | Yiheng Li1 Heyang Jiang1,2 Akio Kodaira1 Masayoshi Tomizuka1 Kurt Keutzer1 Chenfeng Xu1 1University of California, Berkeley 2Tsinghua University |
| Pseudocode | Yes | The algorithm is shown below: Algorithm 1 Batch-wise Image-Noise Assignment |
| Open Source Code | Yes | The code is available at https://yhli123.github.io/immiscible-diffusion |
| Open Datasets | Yes | We conduct extensive experiments on three common modes: unconditional, conditional, and fine-tuning on three diffusion baselines: Consistency Models, DDIM and Stable Diffusion and three datasets: CIFAR-10, Celeb A and Image Net datasets. |
| Dataset Splits | No | The paper refers to 'training steps' and 'evaluations', but it does not explicitly describe the training, validation, and test dataset splits or how data was partitioned for validation. |
| Hardware Specification | Yes | Table 1: Experiment setting. Model... Devices 4 A6000 8 A800 16 A800 1 A5000 4 A6000 8 A800 4 A6000 |
| Software Dependencies | No | The paper mentions software like 'Scipy' and 'Diffusers of Huggingface team', but it does not provide specific version numbers for these or other key software dependencies required to replicate the experiments. |
| Experiment Setup | Yes | The training hyperparameters are shown in Tab. 1. Unspecified hyperparameters are taken the same as those in their baseline methods original papers. Table 1: Experiment setting. Model... Batch Size 512 1024 2048 256 512 2048 512 Resolution 32 32 64 64 64 64 32 32 256 256 256 256 256 256 |