Reproduce “Moodifier: MLLM-Enhanced Emotion-Driven Image Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiarong Ye, Sharon X. Huang
A framework called Moodifier enables precise, emotion-driven image editing by translating abstract emotional concepts into context-specific visual modifications. This system leverages MoodArchive, a large-scale hierarchical emotional image dataset, and MoodifyCLIP, an emotion-specific vision-language model, to preserve image content and identity during transformations.
Key insights
- Detailed, hierarchical emotional annotations, especially when generated by MLLMs and human-validated, provide the granular context required for models to accurately understand and translate abstract emotions into concrete visual attributes.
- A multi-faceted training approach combining global contrastive, fine-grained regional alignment, and optimal transport losses significantly enhances a VLM's ability to reason about nuanced emotions and their spatial manifestations.
- Effective and precise emotion-driven image editing is achievable without retraining diffusion models by strategically leveraging MLLM-generated contextual guidance and attention masks to control the denoising process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.