ParaSolver: A Hierarchical Parallel Integral Solver for Diffusion Models
Authors: Jianrong Lu, Zhiyu Zhu, Junhui Hou
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that Para Solver achieves up to 12.1 speedup in terms of wall-clock time. The results of these experiments demonstrate that Para Solver enhances the efficiency of sequential sampling methods by approximately 2ˆ 12ˆ times, all while maintaining consistent sample quality as measured by metrics like FID score or CLIP score. |
| Researcher Affiliation | Academia | Jianrong Lu, Zhiyu Zhu, and Junhui Hou 1Department of Computer Science, City University of Hong Kong, Hong Kong SAR EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 details the complete process of the proposed Para Solver. |
| Open Source Code | Yes | The source code is publicly available at https://github.com/Jianrong-Lu/Para Solver.git. |
| Open Datasets | Yes | We evaluate Para Solver across various high-dimensional image generation models such as latent-space diffusion model Stable Diffusion-v2 (Rombach et al., 2022) and pixel-space diffusion model LSUN Church (Yu et al., 2015). ...we evaluate FID and CLIP scores on a random selection of 5000 samples from the Image Net-1k dataset. ... For pixel-space models, we employ models pretrained on the LSUN Church dataset. |
| Dataset Splits | Yes | For latent-space models, we leverage the Stable Diffusion-v2 model model and evaluate FID and CLIP scores on a random selection of 5000 samples from the Image Net-1k dataset. ... We compute the FID score on 5000 random samples of LSUN Church datasets. |
| Hardware Specification | Yes | We assess the performance of all methods on 8 NVIDIA RTX 3090 GPUs, each equipped with 24268 MB of memory. |
| Software Dependencies | No | The paper mentions 'Diffusers library' as used for implementation, but does not provide specific version numbers for this or any other software component (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | we apply our Para Solver and Para Di GMS to DDPM with 1000 sequential sampling steps. For DDIM and DPMSolver, we consider two settings: 25 and 50 sequential sampling steps... We tune the best tolerance for Para Di GMS via grid search on t0.001, 0.005, 0.01, 0.5, 0.1u... For our Para Solver on Stable Diffusion-v2 model, we set the tolerance as 0.55, 0.05, and 0.05 for Para Solver on DDPM, DDIM, and DPMSolver respectively. ... For our Para Solver, we set the number of subintervals as 100 and the preconditioning steps as 2. ...with classifier guidance w = 7.5. |