Reproduce “SplitFlow: Flow Decomposition for Inversion-Free Text-to-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
Sung-Hoon Yoon, Minghan Li, Gaspard Beaudouin, Congcong Wen, Muhammad Rafay Azhar, Mengyu Wang
SplitFlow introduces an inversion-free text-to-image editing framework that decomposes complex editing tasks into simpler sub-flows, guided by an LLM, and then intelligently aggregates these sub-flows to achieve precise and disentangled edits. This approach improves semantic fidelity and background preservation, outperforming baselines like FlowEdit and FTEdit on the PIE-Bench benchmark across various metrics.
Key insights
- Decomposing complex editing tasks into semantically distinct sub-flows, followed by intelligent aggregation, effectively mitigates gradient entanglement and improves control over individual attributes.
- Latent Trajectory Projection (LTP) and Velocity Field Aggregation (VFA) can resolve conflicts between sub-flows and enhance flow diversity, leading to a better balance between image fidelity and editability.
- The proposed framework is robust across different LLMs and backbone generative models, demonstrating the generalizability of flow decomposition for controllable image editing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.