Reproduce “InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators” 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

Xuanlang Dai, Jiazi Bu, Yujie Zhou, Yuhong Liu, Beichen Zhang, Penghui Yang, Ziyu Liu, Yuhang Zang

InstantFusion bridges intermediate states between frozen text-to-image generators with different latent spaces. Each generator's time-conditioned Latent AutoEncoder maps states into a shared representation; training aligns states, preserves information through reconstruction, and aligns denoising updates. The Qwen-Image/SD3 hybrid scores 35.06 on GenEval-v2 versus SD3's 22.14, and 0.6488 on CVTG-2K versus 0.3682. The formulation assumes a shared VAE and compatible time parameterization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.