TextLift: Lifting Editable Text Layers from Generated Poster Images for Multilingual Design

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "TextLift: Lifting Editable Text Layers from Generated Poster Images 
for Multilingual Design"

by

KANG Zhaowei

Abstract:

Automated graphic design generation has attracted increasing attention, yet 
existing methods still struggle with text rendering accuracy, editability, 
and multilingual support, especially for long-text poster design. This report 
presents TextLift, a generate-then-decompose framework for producing fully 
editable Chinese and English poster designs from short user prompts. Instead 
of relying on a generative model to render final typography directly, 
TextLift transforms a generated draft into a clean background and structured 
text layers, and re-renders the recovered text deterministically. By 
decoupling text rendering from image synthesis, the framework preserves the 
overall visual style and composition of the draft while restoring accurate 
and editable typography. The output includes a background layer and editable 
text elements with content, position, font, and color attributes, enabling 
post-generation modification that conventional pixel-only methods cannot 
support. Experimental results show that TextLift substantially improves text 
fidelity while retaining most of the perceptual quality of the original 
draft. In addition, the final posters are judged to exhibit stronger 
text-visual integration and a more professional overall appearance than the 
raw generated images. These findings suggest that TextLift provides a 
practical solution for editable multilingual poster generation.

Date            : 29 April 2026 (Wednesday)

Time            : 14:50 - 15:30

Venue           : Room 2126A (near Lift 19), HKUST

Advisor         : Dr. CHEN Qifeng

2nd Reader      : Prof. SANDER Pedro