Idempotence and Perceptual Image Compression
Authors: Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, Ya-Qin Zhang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we show that our proposed approach outperforms state-of-the-art methods such as Hi Fi C (Mentzer et al., 2020) and ILLM (Muckley et al., 2023), in terms of Fr echet Inception Distance (FID). The source code is provided in https://github.com/tongdaxu/ Idempotence-and-Perceptual-Image-Compression. |
| Researcher Affiliation | Collaboration | Tongda Xu1,2, Ziran Zhu1,3, Dailan He4,5, Yanghao Li1,2, Lina Guo4, Yuanyuan Wang4 Zhe Wang1,2, Hongwei Qin4, Yan Wang1 , Jingjing Liu1,6 & Ya-Qin Zhang1,2,6 1Institute for AI Industry Research, Tsinghua University 2Department of Computer Science and Technology, Tsinghua University 3Institute of Software, Chinese Academy of Sciences, 4Sense Time Research 5The Chinese University of Hong Kong, 6School of Vehicle and Mobility, Tsinghua University |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | The source code is provided in https://github.com/tongdaxu/ Idempotence-and-Perceptual-Image-Compression. |
| Open Datasets | Yes | Following previous works in unconditional image generation (Karras et al., 2019; Ho et al., 2020), we train our unconditional generative models on FFHQ (Karras et al., 2019) and Image Net (Deng et al., 2009) dataset. ... To test the generalization ability of our method on other datasets, we also use first 1000 images of COCO (Lin et al., 2014) validation split and CLIC 2020 (Toderici et al., 2020) test split as additional test set. |
| Dataset Splits | Yes | As (Chung et al., 2022a), we split the first 1000 images of FFHQ as test set and the rest for training. And we use first 1000 images of Image Net validation split as test set and use the Image Net training split for training. |
| Hardware Specification | Yes | All the experiments are implemented in Pytorch, and run in a computer with AMD EPYC 7742 CPU and Nvidia A100 GPU. |
| Software Dependencies | No | The paper mentions software like Pytorch and CompressAI, but does not provide specific version numbers for these software components. For example, it states: "All the experiments are implemented in Pytorch..." and "For Hyper (Ball e et al., 2018), we use the pre-trained model by Compress AI (B egaint et al., 2020)". |
| Experiment Setup | Yes | For PULSE (Menon et al., 2020) inversion of Styple GAN, we follow the original paper to run spherical gradient descent with learning rate 0.4. We run gradient ascent for 500 steps. For ILO (Daras et al., 2021) inversion of Style GAN, we follow the original paper to run spherical gradient descent on 4 different layers of Style GAN with learning rate 0.4. We run gradient ascent for 200, 200, 100, 100 steps for each layer. For MCG (Chung et al., 2022b) and DPS (Chung et al., 2022a), we follow the original paper to run gradient descent for 1000 steps, with scale parameter ζ increasing as bitrate increases. ... For five bitrate of Hyper based model, we select ζ = {0.3, 0.6, 1.2, 1.6, 1.6} on FFHQ, {0.3, 0.6, 0.6, 1.2, 1.2} on Image Net, {0.6, 0.6, 0.6, 1.2, 1.2} on COCO and {0.45, 0.9, 0.9, 1.2, 1.6}. |