ReDi: Efficient Learning-Free Diffusion Inference via Trajectory Retrieval

Authors: Kexun Zhang, Xianjun Yang, William Yang Wang, Lei Li

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments demonstrate that REDI improves the model inference efficiency by 2 speedup. Furthermore, REDI is able to generalize well in zero-shot cross-domain image genreation such as image stylization.
Researcher Affiliation Academia Kexun Zhang 1 Xianjun Yang 1 William Yang Wang 1 Lei Li 1 1Department of Computer Science, University of California, Santa Barbara. Correspondence to: Kexun Zhang <kexun@ucsb.edu>.
Pseudocode Yes Algorithm 1 REDI Knowledge Base Construction
Open Source Code Yes The code and demo for REDI is available at https://github.com/ zkx06111/Re Diffusion.
Open Datasets Yes We build the knowledge base B upon MS-COCO (Lin et al., 2014) training set (with 82k imagecaption pairs)
Dataset Splits Yes We build the knowledge base B upon MS-COCO (Lin et al., 2014) training set (with 82k imagecaption pairs) and evaluate the generation quality on 4k random samples from the MS-COCO validation set.
Hardware Specification Yes For example, the basic sampler takes 336 seconds on av-erage to run on an NVIDIA 1080Ti
Software Dependencies No The paper mentions that 'Our inference code is based on Huggingface Diffusers', but it does not specify version numbers for this or any other software dependencies.
Experiment Setup Yes For PNDM, we generate 50-sample trajectories to build the knowledge base. We choose the key step k to be 40, making the first 10 samples in the trajectory the key, and alternate the value step v from 30 to 10. For DPM-solver, we choose the length of the trajectory to be 20 and conduct experiments on k = 5, v = 8/10/13/15. in this experiment, we choose K = 47 and V = 40.