Vorch-IR Long-Form Unified Multimodal Identity Replacement Video Generation

Yaole Wang1*, Xiaoyu Chen2*, Xin Ma1*, Yang Ding1, Gang Yue1, Jingjing Chen2, Lin Ma†, Yaohui Wang1†

1 Vorch Team    2 Fudan University

*Equal contribution †Corresponding authors

Paper Code (Coming Soon)
Overview

Identity replacement aims to replace the identity of subjects in a video while preserving the original motion, expressions, and temporal consistency. Vorch-IR supports single-person and dual-person replacement, optional background replacement, and minute-scale generation within a unified image-text-video editing framework.

01

Independent Subject References

Each subject can be replaced from its own separate reference image.

02

Subject + Scene Replacement

Multiple identities and background references can be fully decoupled.

03

Minute-Scale Inference

Long-form generation preserves identity, motion, and temporal coherence.

Multi-Subject Replacement

Replace multiple people and the scene together while keeping each identity reference independent.

Task 01

Independent references

Dual-subject identity replacement with optional background control.

Long-Form Replacement

Extend identity replacement to long sequences while preserving motion and visual coherence.

Task 02

Minute-long inference

Long-form cases with stable identity over extended video clips.