Reproduce “ViLFlow: Parallel Vision-Language Generation with Continuous Latent Flows” 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
Kevin Zhai, Bian Sun, Zhenyi Wang, Song Wang
ViLFlow generates image-conditioned answers and captions by jointly refining continuous text embeddings over iterative denoising steps, then decoding all positions into words in parallel. An image-guided teacher trains the denoiser with one image-conditioned prediction per step, while a separate decoder learns from denoiser predictions. It reaches 63.39% VQAv2 soft-answer accuracy and 111.16 COCO CIDEr, but greedy autoregression scores higher on CIDEr.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.