RetouchGPT: LLM-based Interactive High-Fidelity Face Retouching via Imperfection Prompting

Authors: Wen Xue, Chun Ding, Ruotao Xu, Si Wu, Yong Xu, Hau-San Wong

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments have been performed to verify effectiveness of our design elements and demonstrate that Retouch GPT is a useful tool for interactive face retouching and achieves superior performance over state-of-the-arts. We utilized Flickr-Face-HQ-Retouching dataset (FFHQR) (Shafaei, Little, and Schmidt 2021) (contains 56k/7k/7k train/evaluate/test images) for comparison.
Researcher Affiliation Collaboration 1School of Computer Science and Engineering, South China University of Technology 2Institute of Super Robotics (Huangpu) 3Department of Computer Science, City University of Hong Kong EMAIL, EMAIL EMAIL, EMAIL
Pseudocode No The paper includes a workflow diagram (Figure 2) and mathematical formulations, but no explicit section or block labeled as 'Pseudocode' or 'Algorithm' with structured steps formatted like code.
Open Source Code No The paper states: 'All competing methods are implemented using open-source codes.' but does not explicitly provide a statement or a link for the source code of Retouch GPT itself.
Open Datasets Yes We utilized Flickr-Face-HQ-Retouching dataset (FFHQR) (Shafaei, Little, and Schmidt 2021) (contains 56k/7k/7k train/evaluate/test images) for comparison.
Dataset Splits Yes We utilized Flickr-Face-HQ-Retouching dataset (FFHQR) (Shafaei, Little, and Schmidt 2021) (contains 56k/7k/7k train/evaluate/test images) for comparison.
Hardware Specification Yes We implement Retouch GPT by using Py Torch and train it on a single GPU with 80G graphics memory.
Software Dependencies No We implement Retouch GPT by using Py Torch and train it on a single GPU with 80G graphics memory. We use the pre-trained T5 model (Raffel et al. 2020) as text encoder and Llama (Touvron et al. 2023b) as LLM. Specific version numbers for PyTorch or the used models/libraries are not provided.
Experiment Setup Yes In the training process, the parameters of Retouch GPT are updated by the Adam optimizer (Kingma and Ba 2015) with the learning rate of 2 10 4. The hyper-parameter κ, ζ, η are set to 10, 0.1, and 0.9, respectively. ... There are a total of 400k training iterations, and the batch size is set to 1.