UniHIR

Draft, Verify, Restore: Self-Refining Historical Inscription Restoration with a Unified MLLM

English | 简体中文

Yuyi Zhang^* · Junle Liu^* · Peirong Zhang^* · Jianliang Liu · Zhenhua Yang · Lianwen Jin^†

South China University of Technology · Deep Learning and Vision Computing Lab

UniHIR is the first unified multimodal large language model for end-to-end historical inscription restoration. It drafts damage locations and illegible content, verifies and refines its own predictions, and restores the full-page appearance in one unified framework.

🔥 News

• 2026-07-05 — The paper is published at ACL 2026 Main. [Paper] [PDF]
• 2026-07-05 — Inference code and pretrained UniHIR weights are released.

✨ Overview

UniHIR replaces task-separated restoration pipelines with a single model that performs three connected stages:

1. **Draft-Guided…