PromptFix: You Prompt and We Fix the Photo

Authors: yongsheng yu, Ziyun Zeng, Hang Hua, Jianlong Fu, Jiebo Luo

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that Prompt Fix outperforms previous methods in various image-processing tasks.
Researcher Affiliation Collaboration 1University of Rochester, 2Microsoft Research
Pseudocode Yes Algorithm 1 High-frequency Guidance Sampling.
Open Source Code Yes The dataset and code are available at https://www.yongshengyu.com/Prompt Fix-Page.
Open Datasets Yes The dataset and code are available at https://www.yongshengyu.com/Prompt Fix-Page. The dataset is available at https://huggingface.co/datasets/yeates/ Promptfix Data.
Dataset Splits Yes For the test set, we randomly select 300 image pairs for each task. We construct the validation dataset with 200 images, and each image contains 3 restoration tasks...
Hardware Specification Yes We train Prompt Fix for 46 epochs on 32 NVIDIA V100 GPUs
Software Dependencies No The paper mentions specific backbone models like Stable Diffusion 1.5 and Intern VL2 [15], but does not provide version numbers for general software dependencies such as Python, PyTorch, or CUDA.
Experiment Setup Yes We train Prompt Fix for 46 epochs on 32 NVIDIA V100 GPUs, employing a learning rate of 1 10 4 with the Adam optimizer. The training input resolution is set to 512 512... we randomly drop the input image latent, instruction, and auxiliary prompt with a probability of 0.075 during training. The hyperparameter λ for the time-scale weight in Algorithm 1 is empirically set to 0.001.