Guiding Instruction-based Image Editing via Multimodal Large Language Models

Authors: Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, Zhe Gan

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate various aspects of Photoshop-style modification, global photo optimization, and local editing. Extensive experimental results demonstrate that expressive instructions are crucial to instruction-based image editing, and our MGIE can lead to a notable improvement in automatic metrics and human evaluation while maintaining competitive inference efficiency.
Researcher Affiliation Collaboration Tsu-Jui Fu1, Wenze Hu2, Xianzhi Du2, William Yang Wang1, Yinfei Yang2, Zhe Gan2 1UC Santa Barbara, 2Apple
Pseudocode Yes Algorithm 1 MLLM-Guided Image Editing
Open Source Code No The paper mentions a "Project website: https://mllm-ie.github.io" but does not explicitly state that the source code for the methodology is available there, nor does it provide a direct link to a code repository.
Open Datasets Yes We use IPr2Pr (Brooks et al., 2023) as our pre-training data. [...] For a comprehensive evaluation, we consider various editing aspects. EVR (Tan et al., 2019), GIER (Shi et al., 2020), MA5k (Shi et al., 2022), and Magic Brush (Zhang et al., 2023a).
Dataset Splits Yes We treat the same training/validation/testing split as the original settings.
Hardware Specification Yes All experiments are conducted in Py Torch (Paszke et al., 2017) on 8 A100 GPUs.
Software Dependencies No The paper states that experiments are conducted in "Py Torch (Paszke et al., 2017)" but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes The learning rates of the MLLM and F are 5e-4 and 1e-4, respectively. All experiments are conducted in Py Torch (Paszke et al., 2017) on 8 A100 GPUs. We adopt Adam W (Loshchilov & Hutter, 2019) with the batch size of 128 to optimize MGIE. [...] During inference, we use V = 1.5 and X = 7.5.