On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation
Authors: Guoqiang Zhang, Kenta Niwa, W. Bastiaan Kleijn
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that considerably better FID scores can be achieved by using IIA-DDIM, IIA-DPM-Solver++, and IIA-EDM than the original counterparts when the neural function evaluation (NFE) is small (i.e., less than 25). |
| Researcher Affiliation | Collaboration | Guoqiang Zhang Dept. of Computer Science University of Exeter United Kingdom g.z.zhang@exeter.ac.uk Kenta Niwa Communication Science Labs, NTT Japan kenta.niwa@ntt.com W. Bastiaan Kleijn School of ECS Victoria Univ. of Wellington New Zealand bastiaan.kleijn@vuw.ac.nz |
| Pseudocode | Yes | Algorithm 1 IIA-EDM as an extension of EDM in Karras et al. (2022) |
| Open Source Code | No | The paper mentions 'the EDM official open-source repository' (https://github.com/NVlabs/edm) which refers to a third-party implementation they used, not their own source code for IIA-DDIM, IIA-DPM-Solver++, or IIA-EDM. |
| Open Datasets | Yes | In this experiment, we tested four pre-trained models for four datasets: CIFAR10, FFHQ, AFHQV2, and Image Net64 (see Table 2 in Appendix D.1). ... using the validation set of COCO2014 over Stable Diffusion V2. |
| Dataset Splits | No | The paper mentions 'using the validation set of COCO2014' and states 'The set-size |B| for computing the optimal stepsizes... was |B| = 200', but it does not provide specific percentages, sample counts, or explicit instructions for how the overall datasets were split into training, validation, and test sets for all experiments. |
| Hardware Specification | Yes | The GPU (NVIDIA RTX 2080Ti) was utilized for measuring the processing time (in seconds). |
| Software Dependencies | No | The paper mentions 'Stable Diffusion V2' and pre-trained models, but does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The set-size |B| for computing the optimal stepsizes when employing the IIA techniques was |B| = 200, which is also the default minibatch size for sampling in the EDM official open-source repository. ... The parameter M in IIA-DDIM was set to M = 10... The set-size |B| for approximating the expectation operation in (7) was set to 16, which is also the mini-batch size for sampling in the computation of the FID scores. The hyper-parameter M in (7) was set to M = 3. |