TransText: Alpha-as-RGB Representation for Transparent Text Animation

Fei Zhang ^1,2, Zijian Zhou ³, Bohao Tang^1,2, Sen He³, Hang Li³, Zhe Wang³, Soubhik Sanyal³, Pengfei Liu^1,2, Viktar Atliha³, Tao Xiang³, Frost Xu³, Semih Gunel³

¹ Shanghai Jiao Tong University     ² Shanghai Innovation Institute     ³ Meta AI

The training and inference code will be released once it has been organized. Please stay tuned.

Project Overview

TransText is the first framework for reference-to-video (I2V) glyph generation. Our framework jointly models appearance and transparency by embedding the alpha channel into an RGB-compatible latent space, utilizing a specialized $\alpha$-reconstruction loss and structural trimap conditioning to ensure high-fidelity, cross-modal consistency in transparent text animations.

Citation

If you find TransText useful for your research, please cite our paper: