Diffusion-based Image Translation using disentangled style and content representation

Authors: Gihyun Kwon, Jong Chul Ye

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results show that the proposed method outperforms state-of-the-art baseline models in both text-guided and image-guided translation tasks.
Researcher Affiliation Academia Gihyun Kwon1, Jong Chul Ye2,1 Department of Bio and Brain Engineering1, Kim Jaechul Graduate School of AI2, KAIST cyclomon,jong.ye@kaist.ac.kr
Pseudocode Yes A.4 ALGORITHM For detailed explanation, we include Algorithm of our proposed image translation mathods in Algorithm 1.
Open Source Code Yes Our detailed implementation can be found in our official Git Hub repository.1
Open Datasets Yes All experiments were performed using unconditional score model pre-trained with Imagenet 256 256 resolution datasets (Dhariwal & Nichol (2021)). For our quantitative results using text-guided image translation, we used two different datasets Animals and Landscapes.
Dataset Splits No The paper does not specify explicit train/validation/test dataset splits with percentages or counts, nor does it refer to predefined splits for training and validation purposes.
Hardware Specification Yes The generation process takes 40 seconds per image in single RTX 3090 unit. All experiments are conducted with single RTX3090 GPU, on the same hardware and software environment.
Software Dependencies No The paper mentions using pre-trained models and referring to official source code of other methods but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes In all the experiments, we used diffusion step of T = 60 and the resampling repetition of N = 10; therefore, the total of 70 diffusion reverse steps are used. For hyperparameters, we use λ1 = 200, λ2 = 100, λ3 = 2000, λ4 = 1000, λ5 = 200. For imageguided translation, we set λmse = 1.5. For our CLIP loss, we set λs = 0.4 and λi = 0.2.