Reproduce “From Tokens to Trajectories: Parameter-Efficient Continuous Adaptation of Unified Multimodal Models” 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
Gayatri Deshmukh, Onkar Susladkar, Adheesh Juvekar, Halil Kilicoglu, Ismini Lourentzou
FlowShift adapts pretrained unified multimodal models for continuous image generation while keeping their backbones frozen. It adds trainable visual interfaces, timestep conditioning, and attention LoRA adapters, using image-based flow matching and teacher-velocity guidance. On UniDFlow-4B, COCO-30K FID improves from 13.33 to 8.92, while GenEval shifts from 0.95 to 0.92; results do not establish gains on every measure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.